CA1281427C - Data processor performance advisor - Google Patents

Abstract

ABSTRACT OF THE INVENTION
A method for evaluating the performance of a data pro-cessor system includes data collection, analysis, and report generation. The data collection involves the periodic se-lection of a number of metrics, which are measurable values in the data processing system, and system parameters, which are system values that can be adjusted. The analysis in-volves the application of certain rules comparing metrics and parameters to corresponding thresholds. The report gen-eration includes the display of recommendations to be taken when certain rules have triggered a certain number of times, and, if desired, the presentation of displays of certain work load characteristics of either the data processor sys-tem by itself or of a network of such data processor system.

Description

I. BAC~&ROUND OF TH~ INVENTION ~ 4~7 The present invention relates in general to the field of automated performance analysis and in particular to automated performance analysis of data processors or data processor net-works using expert systems.
Conventional data processors of more than moderate complex-ity contain fairly sophisticated operating systems, especially if those processors are to have g~neral application. One major pur-pose of an operating system is to maintain control of both a data processor itself and of an entire system containing that data processor. For example, an operating system can control th~ num-ber of processes (i.e., users) that can be resident in the system and control the size of memory allocated to each process. The ; operating system may also control the input and output to exter-15 ! nal devices and to secondary memory, such as disks.
An operating system often exercises its control by adjusting:
certain parameters, such as maximum working storage size allottedi to each process, prioFity of a process, size of cache memory or ' size of common area of memory. The term .~parameters~ is used 20.1 herein to denote system values which can be set either by the operating system itself or by a user. That term is used di~fer-ently from the term "metrics, n which refers to system values which can be measured, such as number of page faults or queue lengths.

4d f~ 2~
The adjustment of parameters affects the operation of the system. For example, the maximum workin9 set size parameter for a process may determine how often that process will have a page fault, or the maximum length of time one process can run continu-ously may affect the amount of swapping by the systemO
In addition, judicious setting of parameters allows an oper-ating system to adapt to different work loads. For example, if a work load consists primarily o~ large batch images (the term "im-ages~ refers to programs run by a "process" or user), the operat-ing system should ideally be configured differently from how itwould be if the work load consists primarily of several small interactive images.
Another reason for "tuning~ an operating system by setting parameters is to adapt to the configuration of a data processor 15 'system or to the configuration of a network containing the data processor system. An operating system should operate differently depending, for example, upon the amount of memory in a processor system and the number and typ~ of I/0 devices in that system.
To determine whether a particular operating system is per-20 , forming efficiently, it is necessary to analyze certain work loadcharacteristics such as the amount of pag;ng, swapping or free memory, the number of computable processes, or the degree o~ disk utilization. Such analysis is often made according to the indi-vidual subjective standards of particular engineers performing the analysis. It would be preferable to make measurements (i.e., 12~3~427 determine the values of metrics) which could objectively indicate an operating system's performance. This has proven difficult, however, because operating system efficiency is not an easy value to measure. Often the data available for evaluation is not a precise measurement of efficiency, but rather only a secondary or tertiary indication.
Many operating systems have built-in programs for making op-erational measurements of a pr~cessor and its operating system.
~ For example, in the VMS family of operating systems for the VAX~
computer systems manufactured by Digital Equipment Corporation, a Monitor program is available to measure pre-ordained operating system data. Such data includes working storage sizes, certain queue lengths, number of disk I/Os, etc. Even with the informa-tion provided by the Monitor program, however, it is still diffi-cult to assess operating system efficiency accurately and to make the proper recommendations. For example, even after determining from the measured data that there is excessive paging in a data processing system, the cause of the excessive paging must be iso-lated before helpful suggestions can be offered.
20 ~ This lack of definitive measurement has been a problem for conventional operating systems even where manuals are provided suggesting ways to improve the operating system performance. For example, associated with the VMS operating systems are manuals suggesting different methods of analysis of that operating sys-tem. One such manual is the Guide to VAX/VMS Performance 0-~ ~s ~

Management, Version 4.4 (April 1986), especially chapter ~. That chapter contains several diagrams which generally ex~lain what conditions to check for, such as excessive swapping or paging.
On-e limitation of that manual, however, is that the analysis called for in the diagrams often depends upon the individual judgments of the engineers performing the analysis. Thus, dif-ferent engineers presented with the same set of facts may have different opinions of, for example, whether there is too much swapp 1 ng .
Variation in the results of operating system analyses by different engineers may occur for several reasons. First, it is not always clear what the diferent measurements representO In addition, different engineers may interpret the measurements dif-ferently. There is also disagreement regarding the weight to be given to the different measurements, thereby resulting in a "seat-of-the-pants" operation by each engineer. Furthermore, since each engineer operates differently, so results oP such ! analyses are nonuniform and generally cannot he reproduced.
Another signficant disadvantage of conventional systems which call for a system engineer's assessment of data is that such systems are generally unhelpful to nonexperts. In addition, even persons with some knowled~e of the operating system may find the manuals of limited help since the manuals often do not gener-ally explain the reasons for taking the suggested actions.

There has been at least one attempt to adapt the technology of artificial intelligence or expert systems to the-problems of performance analysis, but that attempt does not address all the problem~ of using the manuals. The TIMM/TUNER, which is described on pages 109-114 of Diqital Review, May 1986, uses cer~
tain measureme~ts from a VAX computer system with a VMS operating system, and presents those measurements to an operator for eva ation. As compared to the VAX~VMS Performance Management Guide, however, the TIMM/TUNER system offers little additional capabili-ties other than automated queries. Furthermore, the TIMM/TUNERis limited to a single node (i.e~, VAX machine~.
It is therefore an objective of the present invention to provide a method of performance tuning and evaluation which mea-sures the data needed to analyze the performance of a data pro-cessing system or network of data processing systems.
It is another objective of the present invention to provide a method of performance tuning and evaluation which analyzes the measured data according to predetermined rules and thresholds and to recommend actions to be taken according to those rules.
20 i It is yet another objective of the present invention to pro-vide a method of performance tuning and evaluation which explains the actions which have been recommended and which supplies the evidence forming the basis of the recommendations.
Additional objectives and advantages of the present inven-tion will be set forth in part in the description which follows .4Z7 and in part will be obvious from that description or may be learned by practice of the invantion. The objectives and advantages of the invention may be realized and obtained by the methods particularly pointed out in the appended claims.
II. SUMMARY OF THE INVENTION
The present invention overcomes the problems of conventional systems and achieves the objectives listed above first by periodically collec~ing from a data processor system or network of such systems a plurality of metrics and parameters, then by applying rules to those metrics and parameters to see whether the rules are triggered, and last by advising certain actions to be taken when the rules have been triggered a sufficient number of times.
In a broad aspect, the pres~nt invention relates to a method of operaking a data processor system to evaluate the data processor system's performance, the data processor system having a plurality of work load characteristics indicative of the data processor system's performance and also having a plurality of system parameters defining allowable operating ranges for the data processor system, the system parameters being capable of adjustment to affect the data processor system's performance and the evaluation being performed relative to a set of rules specifying acceptable states of the work load characteristics, the method comprising the steps of:
2S measuring, at least once during each of a plurality of predetermined time periods constituting major intervals, a value of each of a plurality of metrics corresponding to saidset of rules, each of said metrics representing a measurable quantity in said data processor system, and said major interval being initiated periodically by said data processor system; storing, in a memory of said data processor system, the value measured for each of said metrics and the major interval during which that measurement was made; comparing the value stored for each of said measured metrics with a corresponding threshold for that metric stored in said data processor system memory, said comparisons taking place for selected ones of said major in~ervals; evaluating the results of each of said comparisons between said stored values and corresponding thresholds to determine whether a plurality of predefined specific relationships for each oP the metrics represented by said stored values and the corresponding thresholds for those metrics have been met, each of said specified relationships relating to one or more of said set of rules; triggering the ones of said rules for which all of the related specified relationships have been met; recording in said data pxocessor memory data representing each of the rulss which have been triggered and the ones of the major intervals during which each of those rules was triggered; counting the number of said major intervals during which each of said rules has triggered, to obtain a total number oP trigyering major intervals for each said set of rules; and recommendlng, from the rules which have triggered, the adjustment of certain ones of said system parameters to improve the performance of said data processor system, said recommending step including the substeps of examining a lisk of suggested actions stored in said memory to find the ones of said suggested actions for which the corresponding rules are rules which have triggered, said list of suggested actions including an identification of the ones of said rules corresponding to each of said suggested actions; ensuring that the total number of said triggering major intervals for each of said rules exceeds a corresponding predetermined threshold number prior to displaying the suggested actions found in said examining step; and displaying the ones of said suggested actions found in said examining step.
In another broad aspect, the present invention relates to a method of operating a data processor system to evaluate the data processor system's performance, the data processor system including a plurality of inter connected elements and having a plurality of work load characteristics indicative of the data processor system'~ performance and also having a plurality of system parameters defining allowable operating ranges for the data processor system, the system parameters being capable o~
adjustment to affect the data processor system's performance and the evaluation ~eing performed relative to a set of rules specifying acceptable states of the work load characteristics, the method comprising the steps of: measuring, at least once during each of a plurality o~ predetermined time periods -7a-constituting major intervals, a value of each of a pluralityof metrics corresponding to said set of rules, each of said metrics representing a measurable quantity in said data processor system, and said major interval being initiated periodically by said data processor system; storing, in a memory of said data processor system, the value measured for each of said metrics and ~he major interval during which that measurement was made, comparing the value stored for each of said measured metrics with a corresponding threshold for that metric stored in said data processor system memory, said comparisons taking place for selected ones of said major intervals; evaluating the results o~ each of said comparisons between said stored values and corresponding thresholds to determine whether a plurality of predefined specific relationships for each of the metrics represented by said stored values and the corresponding thresholds for those metrics have been me~, each of said specified relationships relating to one or more of said set of rules; triggering the ones of said rules for which all of the related specified relationships have been met; recording in said data processor memory data representing each of the rules which haYe been triggered and the ones of the major int~rvals during which each of those rules was triggered; counting the number of said major intervals during which each of said rules has triggered, to obtain a total number of triggering major intervals for each said ~et o~ rules; and recommending, from the rules which -7b-~81427 have triggered changes in a specified configuration of said plurality of interconnected elements to improve the performance of said data processor system, said recommending step including the substeps o~ examining a list of suggested actions stored in said memory to find the ones of said suggested actions for which the corresponding rules are rules which have triggered, said list of suggestion actions including an identi~ication of the ones of said rules corresponding to each of said suggested actions; ensuring that the total number of said triggering major intervals for each of said rules exceeds a*p949Xcorrespon~r~etermined threshold number prior to displaying the suggested actions found in said examining step; and displaying the ones o~ said suggested actions found in said examining step.
~he accompanying drawings which are incorporated and in which constitute a part of this specification/ illustrate an embodiment of the invention and, ~ogether with the description, explain the principles of the invention.
IIIo BRIEF D~SCRIPTION OF TH~ ~RA~INGS
Fig.1 is a drawing of a VAXcluster network o~ VAX
computers, disks and disk controllers which i5 used in explaining the method of this invention;
Fig. 2 is an illustrative diagram of components of a system which can operate the method of this invention;
Fig. 3 is an illustrative digram of the Database shown in Fig. 2;

2 ~
Fig. 4 is an illustrative diagram of the component parts ofthe Advisor Subsystem shown in FigO ~;
Fig~ 5 is a flow diagram of an initialization routine for the data collection procedure Fig. 6 is a flow diagram of a main control loop for the data collection procedure used in accordance with this invention;
Fig. 7 is a flow diagram of a procedure called when measure-ment i5 resuming in accordance- with the main control loop for the data collection procedure whose flow diaqram appears in Fig. 6;
Fig. 8 is a flow diagram of a procedure called when measure ment is being suspended in accordance with the main control loop for the data collection the procedure whose flow diagram is shown:
in Fig. 6 , Figs. 9A~9I contain a decision tree for the memory rules determined in accordance with the present invent;on Figs. lOA and lOB contain a decision tree structure for the . CPU rules determined in accordance with the present invention;
Figs. llA-llD contain a decision tree structure for th~ I/O
I rules determined in accordance with the present invention;
?O ! Fig. 12 is a schematic representation of the storage of transferred data into an evidence table;
.¦ Fiys. 13A-13D are four (~) states of the evidence table shown in Fig. 12 for the transferred data also shown in Fig. 12;
Figs. 14 and 15 are examples of analysis reports generated in accordance with the teachin~s of the preferred embodiment of the present invention;

Figs. 16 and 17 are examples of performance reports gener-ated in accordance with the teachings of the preferred embodiment of the present invention and Fig. 18 is an example of a histogram of CPU utilization gen-erated in accordance with the teachings of the preferred embodi-ment of the present invention.
IVo DE5CRIPTION OF TH~: PRE:FERRED EMBODIMENT
Reference will now be made in detail to a presently pre-ferred embodiment of the invention, an example of which is shown in the accompanying drawing figures.
A! Overview The method of evaluating performance used in this invention is not limited to any specific processing system, but the example of the method illustrated in the accompanying figures will be described with reference to a VAXcluster network of Y~X computers manufactured by Digital Equipment Corporation. The YAX computers are assumed to be using a VMS operating system, version 4.4, supplied by Digital Equipment Corporation.
. Fi~. 1 shows an example of a the VAXcluster network desig-nated ~enerally as 100. In VAXcluster network 100, there are . five nodes comprising VAX computers 110, 120 and 130, and intel-, ligent disk controllers 140 and 150. Disk controllers 1~0 and 150 are preferably models HSC 50 or 70, also manufactured by Digital Equipment Corporation. Although the intelligent disk controllers 140 and 150 are referred to as nodes in the ~ 4 ~7 terminology of VAXcluster networks, the reference to nodes in the description of this invention will, unless noted otherwise, refer only to VAX computers. Disks 1~2, 144, and 146, accessible by the controllers, are preferably conventional disks, and in VAXcluster network 100 shown in Fig. 1, the disks could be RA81 disk drives.
The heart of VAXcluster network 100 is a Star Coupler 160, which is coupled to each of computers 110, 120, 1~0, and to each of disk controllers 140 and 150 via a communications interface ("CI") line. Star Coupler 160 allows interprocessor communica-tion between computers 110, 120, and 130 as well as access of disks 142, 144 and 146 by the VAX computers 110~ 120 or 130.
The evaluation method of this invention is applicable both to data processors, such as VAX computers 110, 120, and 130, and 15 ' to processor networks, such as VAXcluster network 100. The eval-uation of network performance is possible because of the types of data collected and analyzed in accordance with this invention are not generally available with conventional methods of performance evaluation.
20 ~ Fig. 2 shows a diagram of the elements of a system, denoted generally as ~00, which can be used for carryin~ out this inven-tion. System 200 must ~e understood to be merely a representa-tive system. Each of the different subsystems shown need not be separate processors or separate programs~ In Eact ;n the detailed discussion of the preferred embodiment, certain B14'~7 optimizations are made, such as the combination of the Rules Subsystem 270 with Thresholds Subsystem 280. The purpose of showing system 200 in Fig. 2 is to aid in understanding of the invention as a whole.
In system 200, a computer (with terminal) 210 is used to in-itiate computer programs which carry out the method this inven-tion. The system parameters which can he adjusted are for the operating system in computer 210. The elements shown in system 200 include a Data Collection Scheduler 220, which determines when the collection of the necessary measured data will occur, and a Schedule File 230; which contains the times at which the data collection will take place. A Data~Collection Subsystem 2~0 actually collects the data.
The data collected includes a plurality of measurable pro-cess, processor, and network statistics, called ~metrics,~ as well as some current parameter settings. The metrics and parameters collected by Subsystem 240 are placed into a Database 250. Database 250 also includes certain parameter values needed j for evaluat;on..
20,1 An Advisor Subsystem 260 uses the metrics as w211 as certain , system parameters in Database 250 in applying rules in Rules .I Subsystem 270. Those rules often involve a comparison of the metrics and parameters in Database 250 with thresholds in Thresh-olds Subsystem 280.

~'~ 8~2~
When Advisor Subsystem 260 determ;nes that certain of the rules in Rules Subsystem 270 have been triggered, an~ have been triggered a sufficient number of times, then messages correspond-ing to~those rules are generated using message templates in Mes-s sage Templates Subsystem 2900 The messages also include certain other data which are used to form reports in Reports Subsystem 295. Those reports are available for display or printing.
Fig~ 3 is a diagrammatic representation of Database 250, Fig. 3 is also intended merely as a representation of the data-base and not intended to define how such a database must look.Database 250, as shown in Fig. 3, includes several files 300. In each of the files, there is also an idcntification of the node (i.e., computer~ for which the measurements in the file pertain and the date at which the measurements were made.
As shown in Fig. 3, files 300 contain records 310 each with certain time periods of measurement. Those time periods, called major intervals, indicate the frequency with which measurements are taken. For records 310, those major intervals are repre-sented as two-minute intervals, which is a default condition in 20,l the preferred embodiment. Other major intervals may also be cho-sen.
Each of records 310 contain subrecords 320, examples of which are shown also in Fig. 3. As shown in Fig. 3, subrecords 320 includes time stamp subrecords, which contain the time of the measurement, performance subrecords, which contain system wide 1 ~ 81~7 performance metrics, and parameter subrecords, which contain sys-tem wide parameter values.
Subrecords 320 also includes image and process subrecords which describe the images that were activated by each o~ the pro-cesses, such as their working set sizes, fault rates, and usernames. As explained above, the term "processes" refer to users, and the term "images" refer to programs run by processes.
The disk subrecords in subrecords 320 describe the activi-ties in the disks. Such activity could include input/output (nI/O~) rates, I/O sizes, or busy times.
The metrics and parameters which are measured and adjusted in the preferred embodiment will not be separately described since those parameters and metrics are unique to each particular system. Instead, particular metrics and parameters are described in the next section as part of the explanation of the rules.
Fig. 4 shows a representation of the component parts of the ; Advisor 26~. Advisor 260 performs two functions. The first is analysis for tuning (element 410~. To perform such analysis, Advisor 260 compares thresholds in Thresholds Subsection 280 to 20 1l the metrics and parameters in Database ~50 and determines whether the rules in Rules Subsection 270 have been triggered. Based on those rules, Advisor 260 reaches certain conclusions about the performance of system 210. Preferably, such conclusions are reached only after particular rules have been met more than a certain number of times in an analysis session.

~ 4'~
The analysis presented by Advisor 260 is displayed to the user or printed in the form of messages often containing recom-mendations. If the user requests, the "evidence" for the recom-mendations will also be printed out~ Such evidence is in the form of a statement of the rule accompanied by the parameters, thresholds, and metrics which were used to determine that the rule is met.
Advisor 260 can also organize work load characterization data (element 420), both for individual processors and for the network as a whole. Such data is a compilation of information about the processes, such as percentage of central processor unit (nCPUn) utilization, characterization of work load between interactive jobs, batch jobs, overhead, etc., or amount of disk utilization. The particular characterization available for this invention is quite detailed due to the type and amount of data collected.
~ Data Collection Figs. 5-8 show flow charts for a method of data collection in accordance with the present invention. The method illustrated in Figs. 5-8 has been used on a VAXcluster network of with VAX
, computers, model 780, that are using version 4.4 of VMS operating ; system.
i Fig. 5 shows an INITIALIZATION sequence for data collection.
The sequence starts at step 500 and the first question asked is whether the particular operating system is supported by the ... .
. . . .

~ 4Z7 program used for implementation of data collection (s~ep 505).
If not, the system prints an error message (step 515) and the procedure is exited (step 510).
If the operating system is supported by the program, then a parameter for the maximum number of processes (e.g., MAXPROCESSCNT for the VMS operating system) is set to the minimum of either the total number of users or some fixed number, such as 512 for VMS operation (step 52~). The principal reason for setting the maximum number of processor parameter is to allow creation of data structures. The fixed number in step 520 should be adjusted to the particular computer and operating system used.
The next step involves the setting of a scheduling priority level (s~ep 525). In the preferred embodiment of the invention, this level is set to 15. The scheduling priority level must be properly set for data collection so that the data (metrics and parame~ers~ can be collected in as close to rea~ time as possible without interfering with the operation of the system. Thus, the priority level should preferably be set helow the priority of the swapper or below any processes that critically need real time re-20 j sponse, hut the priority should be set above most other pro-cesses.
The next step involves enabling an asynchronous system trap.
(nASTn) to signal to the sytem that a review of the measurement schedule is needed (step 530). The AST is a VMS construct and it is enabled in step 530 so that the system will be alerted when the scheduling of measurements changes. Generally, the scheduling of measurements is kept in a schedule file; The pur-pose of enabling the AST is to advise the system to check the schedule file when a change has been made to that measurement file After enablement of the AST, the data measurement schedules and intervals are read from the schedule file (step 535). The measurement schedul*s indicate-the times of day during which mea-surement should take place, and the intervals indicate how often the measurement shold take place. In the data measurement sched-ules, there is a parameter for scheduled start time and the sys-tem must test whether that start time is in the future (step 540). If it is, then the system hibernate~ until the scheduled start time (step 545).
If the scheduled start time is not in the future, then the scheduled start time is tested to see whether it is in the past (step 550~. If the scheduled start time is in the past, then the;
system hibernates indefinitely or until the schedul~ file is ch~nged (step 555), and the procedure is exited (5tep 560j.
20 j If the scheduled start time is neither in the past nor in the futur~, then the system chooses a timekeeper node and the . cluster logical clock of that node is initialized (step 565). A
timekeeper node is used to synchronize measurement by a network of processors. In a network of processors operating in accor-dance with the preferred embodiment, each processor has its own ~ '7 internal clock which may not be synchronized to the other proces-sors' clocks. Rather than change the internal timing of each one the processors, one processor is chosen to be the timekeeper node and that processor stores a logical cluster clock. The logical cluster clock is related only to measurements in accordance with data collection method of this invention. Each of the other pro-cessors keeps its own logical clock and, as explained in detail below, ensures that that clock-is synchronized with the cluster logical clock of the timekeeper node.
Next, the system initializes all the mea~urements (step 570). This initialization provides reading of the metrics so that incremental values can later be measured during the data measurement loops. Finally, the initialization sequence is exited (step 575).
Fig. 6 shows the main con~rol loop for the data collection.
In the main control loop shown in Fig~ 6, there are ~wo measure-ment intervals. One is a major interval during which most of the system metrics are measured and during which the values are stored. In the preferred embodiment, the default condition for 20 I khis interval is every two minutes.
! There is also a minor interval for measuring certain metrics which change too quickly for accurate representation every major interval. 5uch ~ast-changing metrics often include certain pro-cessor metrics, such working storage size. For such fast-changing metrics, a measurement is made once every minor 8~
interval. In the preferred èmbodiment, the deault condition ~or the minor interval is five seconds. A running average over a major interval is taken of the metrics measured during minor in-tervals.
In the main control loop shown in Fig. 6, the procedure is started (600) and the cluster logical clock (i.e., the logical clock at the timekeeper node) is read (step 605). Each separate processor then recalibrates its own logical clock to the cluster logical clock periodically, such as every 30 minutes (step 610~.
Next, the absolute ~ime of day is calculated, and various timers~
such as the major and mino~ interval timers~ are reset using the logical clock (step 615).
Next, the procedure waits for a minor interval. In addi-tion, if the VMS operating system takes certain action, such as flushing disk measurements (step 620~, then step 620 is also entered so that measurements will only be made at the end of a minor interval.
At the end of a minor interval, the question is asked wheth~
er measurement would b~ resuminq tstep 635). For example, mea 20 i surement would be resuming if measurement had bee~ susp~nded pre-viously but should now 3tart aqain. If measurement is resuming, I then procedure A is followed (step 640). Procedure A is shown in detail in Fig. 7.
As shown in Fig. 7, procedure A is entered tstep 700) and the appropriate daily database file is either created, if it does ~ '7 not already exist, or appendèd, if it already does exist (step 710). Next, the expired database files, for example from previ-ous days, are deleted (step 720).
Next, special measurement code for performance evaluation is loaded into the non-paged pool, if necessary (step 730). The non-paged pool is a common area of memory which can be accessed by any process. Measurement code is code executed by an imageD
to take certain image or process measurements if such measure-ments cannot otherwise be taken. Images can access measurement code because it is in the non-paged pool.
Next, image rundown is instrumented if necessary (step 7~0~.
In this step, the ~IM5 is instructed to execute the special mea-surement code for a process whenever an image is ended. Image rundown involves certain housekeeping functions, such as closing files that have been opened and deallocating memory space.
If necessary, the disk I/O Perform is then instrumented Istep 750). Disk I/V Perform monitors or measures disk input output activity.
Procedure ~ is then exited, and, as shown in Fig. 5, pro-cessing resumes at step 535, i.e., reading data measurementschedules and intervals from the schedule file.
; In Fig. 6, if the test in step 635 shows that ~he measure-ment is not resuming, then a determination must be made whether measurement is suspending due to scheduling (step 645). If it is, then procedure B (step 650) must be executed. Procedure B is shown in detail in Fig. 8.

~ 8~7 As shown in Fig. 8, procedure B is entered (step 800) and the image rundown is then deinstrumented (step 810) Next, the instrumentation of the disk I/O Perform is then removed (step 820). Finally, the daily database file is closed (step 830) and, after waiting until the end of the hour (step 840) the procedure is exited (step 850). After leaving procedure B, the next deter-mination made is whether measurement is resuming (step 635).
In accordance with the main control loop shown in Fig. 6, if measurement is neither resuming nor suspending, then the data measurement takes place. New daily files are created and expired files are deleted if it is midnight (step 660~. Next, the fast changing process data described above is measured and level aver-aged every minor interval (step 665) Preferably the level averaging involves maintaining a running average over a major in-terval time period.
Then, the non-paged database/ which is the database in the non-paged pool containinq measurement information, is flushed (step 670)a The non-paged database includes, for example, infor-, mation measured during image rundown. Flushing the da~a base in-20 l volves rewriting the data into a data collection buffer prefer-I ably in Database 250.
The main control loop shown in Fig. 6 then makes the neces-sary measurements every major interval. These measurements in-clude the parameter values and the remainder of the metrics not measured every minor interval. For example, a time record is ~ 7 recor~ed every major interva~ (5tep 675), as are operating system statistics ~step 680). In addition, parameter value~, disk records and communications records are also recorded every major interval (step 6a3, 686 and 6~8, respectively). The parameter values are, as explained above, those values which are set either by the operating system or by the user. The disk records are metrics related to disk I/O as distinct from the operating system statistics or metrics which are recorded in step 680. The commu-nications records refer to communication resource metrics, and are described in the explanation of the rules belowO
The configuration record is also read at the end of a major interval if the configuration of the network or processor has changed ~step 690)~ For example, if a processor has gone down or a disk has been removed, then the configuration has changed and a lS new record will be read.
Next, the process database in the non-paged pool is then up-dated for each process. Finally, the process disk statistics are recorded at th~ end of a major interval (step 696).
At the end of the main control loop, the recorded statistics 20 j in the non-paged pool are written into a database, such as ¦ Database 250. In addition, that database is checkpointed, which involves updating procedures to inform the entire system that new data has been recorded and to adjust the sizes of the database.
The main control loop then continues at step 620, where the sys-tem waits until the end of a minor interval before continuing.

~ 7 - C. Analysls In general, the analysis portion of this invention involves the application of certain rules to the metrics and parameters collected during the data collection operation. The rules in general involve the comparison of metrics and parameters either to themselves or to certain thresholds to see whether the criteria of the rules (e.g., greater than, less than, etc.) have been satisfied. If all the criteria for a rule have bYen satisfied, then the rule is said to trigger. When a rule trig-gers more than a predetermined number of times, then a recommen-dation, included as part of a message, is made in accordance with the rule. That recommendation generally involves suggestions to the user that adjustment of some parameter or a change in system configuration could be made to improve system performance.
The rules involved in the preferred embodiment can generally be classified into memory rules, CPU rules, I/O rules, channel rules, resource rules, and cluster rules. Presently, due to the number of rules in the preferred embodiment, the implementation of the rules is done by way of programs which incorporate both 20 i the thresholds and ~he rules together. Alternatively, especially , as a system grows, the implementation can be replaced by the use of an inference engine and ~nowledge base of rules which can be more easily modified than can be the programs.
Figs. 9A-9I include a decision tree structure for rules 1-32 which correspond to the memory rules. Figs. lOA and 108 include ~ '7 a decision tree structure for the CPU rules, shown as rules 33-39, and Figs. llA-llD include a decision tree structure for the I/O rules, shown as 40-Sl in the preferred embodiment. In addition, channel rules, resource rules, and cluster rules which do not lend t~emselves to decision tree structures, will instead be described textually.
In the decision tree shown in Figs. 9A-9I, lOA and lOB, and llA-llD, the circular elements-are either decision points or stop points indicating respectively, tests to be made or exits from a particular decision tree. The tests which may be made are printed along side of the circular nodes and are explained in additional detail in the text. Th~ sq~are boxes contain rule numbers. The message template for the rules are in Appendices 1-6 at the end of this description.
15 ! Preferably, one of the proc~ssors in the network tests the rules using all of the metrics and parameters which the proces-sors have stored in a database. Each record is tested sepa-rately, so the rules are first applied to the metrics gathered during the first major interval under consideration. The rules 20 i are then applied tested to the metrics and parameters gathered during the second major interval, and this pattern continues until all of the major intervals in the desired analysis time are covered. Preferably, a system mana~er or engineer can specify to the analysis program for which time periods analysis is desired.
In the preferred embodiment of the method of this invention, a ~ 7 system-will create a data strùcture during the analysis proce-dures indicating the rules which have been fired as well as the evidence for those rules. The evidence includes the values of the metrics and parameters which were tested, as well as the cor-responding thresholds, ~hich gave rise to the triggering of the rule. Those data structures will be described in greater detail in the Report Generation section of this description.
In Fig. 9, the first test-made for the memory rules is at decision point 900. That test determines whether the page fault rate from the disk is greater or equal to some threshold, such as 10, whether the page faults from cache are greater than or equal to another threshold, such as 100, whether the in-swap-rat~ is greater than or equal to another threshold, such as 1, or whether the sum of freecnt + mfycnt is less than or equal to FCEEGOAL +
~ILIMIT (many of the parameters used in this description of an embodiment of the invention use a shortened form for ease of ref-erence; thus "HILIMIT" is properly "MPY_ HILIMIT)o 8asically, the test at decision point 900 is whether there is either too much paging or too much swapping, or whether there is scarce free mem-ory.
;l The test or too much pag;ng involves a comparison of the page fault rates both from disk and from cache to different thresholds, for examples 10 and 100, respectively. Page faults from dis~s are called hard faults since they are generally more costly from a time standpoint than page faults from cache, ~hich ~ 4'~

are called soft faults. The page fault rates from disk and the page fault rate from cache, both of which are in occurrences per second, are metrics measured during the data collection opera-tion.
The question of too much swapping is determined by the in-swap-rate, which determines whether any programs have been swapped out. The in-swap~rate, which i.s also a metric, and mea-sures the number o times that processes have been swapped.
Scarce-free memory is determined by the remaining test shown in decision point 900. FREEGOA~ and HILIMIT are, respectively, the limits for the number of page5 of free memory and the number of pa~es on the modified page list which are available in memory.
A page on the modified page list for a process that is no longer active, but such a page contains changes that must be writ~en back onto disks. The freecnt and mfyont metrics refer to the actual number of pages on the free list and the actual number of pages on the modified page list. Those two parameters together represent the amount of free memory, and if they are less than the desired goal for that free memory, then the determination is 20 I that free memory is scarce.
I there is neither too much paginq, too much swapping, or scarce-free memory, then the procedure for testing memory rules reaches stop point 902 stops since ~here has been no memory prob-lem during the corresponding time period.

4~'7 ~ f, however, there has bèen too much paging or swapping, or there is scarce free memory, then the decision node 904 is reached and a determination is made whether an image has had at least a threshold number of page faults (both from disk and from s cache), such as 500 faults, whether its CPU time (a metric indicating how long an image is in execution mode) is greater than or equal to another threshold, such as ~ seconds, and wheth-er that image's up-time (a met-ric indicating the duration that the image exists~ is greater than or equal to yet another thresh-old, such as 30 seconds. If so, then rule 1 triggers. The mes-sage template for that rule (Appendix 1) is then printed out along with information about the name of the image, the name of the user, the time, the image fault rate and the total fault . rate. Basically, rule 1 triggers if an application program is 15 , either a special type or is poorly designed. The parameter dis-cussed in the message WSQUOT~ is the working size quota for a particular process. That parameter may be increased ~o allow more pages to be stored for that imaye, so fewer page faults should occur.
20 1l Decision node 906 is reached whether rule 1 is triggered or not, and the same paging questions as were asked at decision ~¦ point 900, specifically, whether the page fault rate from disk or i the page fault rate from cache too high, are asked again. If so, then decision point 908 is reached.

14'~7 Decision point 908 invoives the comparison of the total image activations to a threshold, for example, .5 per second. I~
the answer is yes, a decision is made that there are too many image activations, and at decision point 910 the question is S asked whether any process ca~ses more than a certain number of activations per second, again, .5 in the preferred embodiment.
If so, it appears that the activations are caused by one process, so rule 2 triggers. The messa~e template for rule 2 (Appendix 1) indica~ees that there are too many image activations causing too many page faults, The suggestion in that message is that a poor-ly written command procedure whlch activates too many programs may be the cause. Pr;nted out along with the message is the time, the user names, and the number of images activated.
If the answer at decision point 910 is no, than rule 3 trig-gers which indicates that there are too many page f~ults associ-ated with too many image activations in the systems as a whole.
As the message template for rule 3 indicates, the cause may be either poorly designed application programs or attempts to run too many command procedures frequently. Along with the message, the time, the total fault rate, the hard fault rate, and the imaqe activation rate are printed out.
I~ the answer at decision point 908 was no since there were not too many image activations, then at decision Point 912 the page faults from disk are again compared to a threshold. If the disk page fault rate is greater than the threshold, indicating 1'~8~4q~7 too hiyh of a hard fault rate, then decision point 914 is reached.
At decision point 914, the total page faults are compared to a threshold, for example 100, and the sum of FREEGOAL and HILIMIT
is compared to the smaller of either 3000 or five percent of the user pages. The purpose of this second determination is to com-pare the page cache, which is the sum FREEGOAL ~ HILIMIT, to either five percent of the mem~ry assigned to the working sets or to 3000. This is a way of testing whether the page cache is suf~
ficiently large. If not, then the decision point 916 is reached.
At decision point 916 the sum of freecnt and mfycnt is again compared to the sum of FREEGOAL and HILIMIT to determine whether free memory is scarce. If not, stop point 91a is reached and the ,process leaves the memory rules.
If the determination at decision po;nt 916 is that free mem-ory is scarce, rule 4 i5 triggered. The associated message, for rule 4, as shown in Appendix 1, indicates that excessive hard faul~ing may have been caused by too small of a page cache, which , can be increased by increasing the parameters specifying page icache size, such as MPW_LOLIMIT, MPW_HILIMIT, MP~_THRESH, FREEGOAL, and FR~ELIM.
! If, at decision point 912, the hard fault rate was not found to be too high, than decision point 920 is reached which compares the total number of page faults to a threshold, for example 100.
This is another way of testiny whether the soft pa~e fault rate ~ 4'~'7 is to~ high. If not, then stop point 922 is reached no more mem-ory rules are tested.
If, however, the soft page fault rate tested at decision point 920 is too high, then decision point 924 is reached, ~hich, as indicated in the explanations of decision points 914 and 916, is a determination of whether the page cache is now too big and free memory is now scarce. If so, rule 5 is triggered. The as~ociated message template in-Appendix 1 contains recommenda-tions for reducing the page cache. Also, the time, total fault rate, image activation rate, hard fault rate, and size of the free memory list tfreecnt) are printed out.
Decision point 926 shown in Fig~ 9C is reached if either decision point 914 is satisfied (high soft fault rate or suffi-ciently large page cache) or if dec;sion point 924 (no large page cache or no scarce memory) ;s not satisf ied. At decision point 926, two determinations are made. The first is to find processes in the low half of the faulting processes, i.e., ones that are not faulting too much, which are not in the COM (compute) mode.
The next determination is whether such processes have a working set size greater than five percent of the usable memory, have page faults less than PFRATH (~he maximum rate which the VMS
operating system assigns for a process), and have a working set size less than or equal to their quota. If those conditions are met, rule 6 triggers. As the message template in Appendix 1 for this rule printed out the triggering of this rule involYes a 14'~7 deter~ination that the total system wide page fault rate was high while certain moderately idle processors had large working sets.
The suggestion is to decrease the working set size quota, WSQUOTA, for certain users, and to decrease that quota for detached processes. The elements printed out with the message or this rule are the user name, the corresponding working set quotas and working set extents, the image name, the time, and the working set size. Also printed out are the image fault rate~ the total fault rate, and the size of the free lis~.
After rule 6 is met, a determination is made at decision polnt 930 whether a particular process has more than twice as many page faults as it~ upper limit and has either a CPU time greater than a certain number, such as .5 seconds, or has its COM
variable, indicating the number of times that process was in the compute mode, more than 1. If those conditions are met and if that process has a working set size greater than WSEXTENT, which is the maximum size of the working set, minus 2/3*WSINC, which is the number of pages that the YMS operating system will add if you need additional space, rule 7 tr;g~ers. Rule 7 also triggers for , the sam~ conditions tested at decision point 928, which is reached when the conditions at decision point 926 are not met.
The determination at decision points 928 and 930 are whether the user working set extents are too low. Rule 7 trigger~ if it appears that the user's working set extents were too low. The conclusion is based on the fact that certain users were running 2~7 images~which wanted more memory than were allowed, and, as ex-plained in the template for rule 7 in Appendix 1, WSEXTENT should be increased for certain users. Also printed out with the mes-sage template are user names, working set quotas and extents, image names, times, working set sizes, image faul~ rates, total fault rates and size of ~he free list.
If the answer at decision point 930 is no, then stop point 931 is reached. If the answer-at decision 928 is no, then deci-sion point 932 is reached which questions whether the ~otal num-ber of processes is greater than ~ALSETCNT. BALSETCNT is a parameter indicating the number of processes that are preconfigured with data structures. In addition, a question is asked whether there is any extra free memory ~freemem~ is equal to (freecnt + mfycnt) minus (FREEGOAL ~ HILIMIT), and whether the lS in-swap-rate was greater than zero. These last two questions are to determine whether there is any-extra free memory and whether there are processes being swapped.
If so, than rule 8 is triggered, indicating that there is excessive page aulting although there is enough free memory.
The message template for rule 8 (Appendix 1) indicates that the , swapper is apparently trimminy the working sets unnecessarily to prepare for a swap. The recommendation in the message template is that the parameter BALSETCNT be increased to prevent the swapper from trimming the processes when there is sufficient mem-ory, thereby reducing the page fault. The data printed out along with the message include time, total fault rate, hard fault rate, image activation rate, size of free list and number of processes in the system.
If the answer at decision point 932 is no, then decision point 934 is reached ~ith compares the BO~ROWLIM and GROWLIM
parameters. The BORROWLIM parameter indicates whether a process qualifies to receive additional memory, and the GROWLIM parameter indicates whether a process will actually receive that memory.
If BORRO~LIM is less than GROWLIN, rule ~ triggers. The solution ~ to the problem indicated ~y rule 9 is suggested in the message template for that rule, and that suggestion is to decrease the G~OWLIM parameter below the ~ORRO~LIM to allow allocation of available memory. In addition to the user name, image name, working set size, image fault rate, the time several parameters ',¦ are also printed out along with the message, including PFRATH, BORROWLIM, GROWLIM, PAGE CACH~, FREELIM, F~E~GOAL, WSMAX, the later parameter indicating the maximum working se~ size.
If the result of the decision that decision point 934 is no, i.e./ BORROWLIM is greater than or equal to GROWLIM, then deci-~O I sion point 936. At decision point 936, WSINC is tested to see if ! it is equal to zero. This state occurs if the A~SA (automatic working set adjustment) is turned off. If so, then rule 10 trig-gers. The problem is that the ~MS cannot allocate additional memory to alleviate the excessive page faulting. This is ex-plained in the message template for rule 10 in App2ndix 1. This .

4'~
problem can be corrected by setting WSINC to 150 or to some other appropriate value.
If the determination at decision point g36 is that WSINC
does not equal zero, then decision point 938 is reached at which several other determinations are made. The first is whether the free memory is greater than some threshold, such as 100. The second is whether there are either two processes or five percent of the processes which have too many page faults (for example more than or equal to 10), have a CPU time which is too large I (such as greater than or equal to .2 seconds), and have a wssiz Iwhich is less than WSEXTEN~ minus WSINC, and a wssiz which is .Igreater than (7/8)*~SLIST. WSLIST is the potential size of the ¦working set. If so, then it appears that the A~SA is too slow .I since the working set size could have been increased to more .Ipages, but was not.

i At that positive determination, decision point 940 is . reached and a determ;nation is made whether there are any users waiting in the RSN$_SWPFILE. That iS the file indicating all of .
the processors that are waiting for memory space but do not fit.
~ 20 ¦ If there are processes in the wait state for the swap file, then .~ rule 11 is triggered. As the template for rule 11, which appears in Appendix 1 points out, increasing the swapping ile will allow.
processes to grow and eliminate faults. The data printed out with the message template includes time, total fault rate, hard . fault rate, image activation rate, size of the free list, and number of processes on the system.

4~
I~ the determination at~decision point 940 is no since pro-cesses are not waiting in the swap file, then PFRAT~ (page fault rate high) is tested against a threshold, such as 160, at deci-sion point 942. If that condition is met, then rule 12 triggers indicating that AWSA is slow to respond since there was excessive page faulting although there was excess free memory. The message template for rule 12 is in Appendix 1~ When the message is printed out, the value for PFRATH is printed out as well as the . time, total fault rate, hard fault rate, image activation rate, and size of the free listo : If the condition and decision point 942 is not met, then decision point 944 is reached at which the size of WSINC is com-pared with a threshold, such as 100. If WSINC is less than that ', threshold, then rule 13 triggers. The message template for rule ¦ 13 is shown in Appendix 1 indicates that the reason for the slow response of the AWSA appears to be that the working sets grow too ¦ slowly, and increasing WSINC may improve this condition. Also ,¦ printed out with the message template would be the time, total ' fault rate, hard fault rate, image activation rate, and size of 1 the free list.
¦ If the result at decision point 9~4 were no, than decision point 946 is reached which compares the parameters AWSTIM to a i threshold, such as 50. The AWSTIM parameter is the amount of CPU
. j time between consecutive working set adjustments. If this number is too high, then rule 14 triggers. The template for that -3~-message indicates that the ~WS~IME parameter, which is the mini-mum time between workinq set adjustments, should decreased such as 220, or the AUTOGEN utility can be allowed to reset AWSTrM.
The ti~e, total fault rate, hard fault rate, image ac~ivation rate, and size of the free list would be printed out or displayed along with the message template. If the result of the determina-tion at decision point 946 is no, then stop point 948 i5 reached.
If the determination at decision point 938 was no and the AWSA was not found to be too slow, then decision point 950 is , reached, which questions whether the voluntary decrementinq has i been turned on. If not, then decision point 952 is reached and two o~her tests are made. The f;rst involves determining whether !
the top two processes which are faulting having smaller working set sizes than the processes with the two largest working set ! size processes. The second determination is whether freecnt is less than BORROWLIM plus WSINC. If both conditions are met, then !
l voluntary decrementing is needed and rule 15 triggers. The mes-i sage template for rule 15 in Appendix l suggests correction by setting parameter WSDEC (the amount by which the working set size i can be decrem~nted) to the appropriate value, such as 35, and by setting the parameter PFRATL (the page fault rate low parameter) to a lower number, such as 10.
If, at decision point 950, it is determined that the volun- ;
tary decrementing has been turned on, than decision point 954 I reached. At that decision point a determination is made whether 4~'~
the page fault rate from dis~ (hard faults) are less than a cer-tain number, such as 10, and whether the top two faulting proces-sors (l) have sizes which are less than WSEXTENT minus WSINC and (2) have a pfrate (the page fault rate per process) greater than PFRATH. If bOth conditions are met, then the AWSA may be shrinking the working sets too much so rule 17 triggers. The template for rule 17 in Appendix 1 indicates.that WSDEC and/or PFRATL should be decreased. Also printed out with the message template would be the time, total fault rate, image activation i rate, hard fault rate and size of free list.
If e;ther t~e results of the determinations at decision points .952 or 954 are no, decision poin~ 956 is reached. Several . conditions are tested at decision point 956~ The first is wheth-l er freemem is too small (e.g., less than or equal too 100). The second is whether SWPOUTPGCNT, the size to which the process is cut back in a swapper trimming, is less than another threshold, . I for example 200. The third condition is whether one-third of thej : , proc-esses are near (for example, plus or minus 30 pages) of theirl ¦ quota or of SWPOVTP~CNT. If all three conditions are met, it appears that several processes have been cut back for no reason, and the swapper timming has been too severe, so rule 16 triggers The template for this rule indicates that an increase in LONGWAIT
. which is th~ length of time that distinguishes idle or abandoned l processes from momentarily inactive processes, would force the swapper to ~ive momentarily inactive processes longer time before they are trimmed. Also printed`out in addition to the template would be time, total fault rate, image activation rate, hard fault rate, average working set size and maximum working set size.
If swapper trimming was not found to be too severe at deci-sion point 956, then freemem is again compared to some threshold, such as lO0, at decision point 958 to see whether free memory is scarce. If not, the stop point 959 is reached. If free memory is scarce rule 18 triggers because it appears that excessive page faulting exists due to heavy memory demand. The message template ' for rule 18 in Appendix l explains this condition in greater ! detail. With that template, the time, total fault rate, image activation rate, hard fault rate and size of the free list would ! be printed out, :I The previous rules have been concerned with paging. If, j¦ however, at decision point 906, too much pasing was not found, i then the question of swappin~ must be investigated, which occurs at decision point 960. In particular the in-swap-rate is com-pared to a threshold, such as 1, at this point. If the in-swap-1 rate is greater than or equal to l, then decision point 962 is .~l i reached.

! At decision point 962~ the free balance set slots are com-pared to another threshold, such as 2. The balance set slots l each correspond to one process that can be coresident. Too small - of a value for the free balance set slots means that there is Z~7 excessive overhead because of swapping and the ~ALSETCNT may be too small since there are no free balance set slots. If the result of the determination at decision point 962 is positive, rule 19 is trlggered. The associated template for this rule is shown in Appendix 1 which would be accompanied by the time, the in-swap-rate, the CPU idle time, the free page ratio, the number of processes in the balance set, and the number of out swapped processes.
If at decision point 962 a sufficient number of free balance .slots were found, then decision point 964, the sum of the ppgcnt plus cpgcnt for all the processes are compared to the total num-~ber of user pages minus the cache minus 100. The metrics ppgcnt and gpgcnt refer to the number of process pages and the number of;
~global pages, respectivelyO The size of the cache is generally ~the size of freecnt plus mfycnt, and the last term, 100, can be .¦changed, but generally refers to some sort of buffer number or ,Islack. The metrics and parameter~ being considered at decision l¦point 964 are used to determine whether there is sufficient memo-, .¦ ry for all of the working sets (i.e., ppgcnt plus gpgcnt exceeds 1 the other term). If so, then stop point 965 is reached. If not, than decision point 966 is reached.
At decision point 966, the size of the cache is compared to . the smaller of either five percent of usable memory or to some other number, such as 3000. If the cache is larger than the smaller of those two numbers, then rule 20 is triggered. The .

~ 8~ ~'7 messa9e template for this rulè appears in Appendix 1 and, when printed or displayed, is accompanied by the time, in-swap-rate, CPU idle time, free balance set slots, number of memory pages availabIe, and the number of memory pages used.
If the determination at decision point 966 is no, then deci-sion point 968 is reached at which a determination is made wheth-er the top two processes together have more than some amount of usable memory, such a5 five pe~cent, and whether they are either in the COM (compute) mode more than half the time or collectively consume more than half of the CPU. If they do, rule 21 triggers j since a determination is made that there are large compute-bound j . processes which are devouring system resources. Several recom-I mendations for this condition are shown in the message template .I for rule 21 in Appendix 1, The data printed out in addition to j the message would include user name, working set quota, image ~I name, working set size, time, in-swap rate, number of free bal-: '. ance set slots, and size of the free listO
If at decision point 968, no large compute-bound processes . were detected then deci~ion point 970 is reached, which looks for¦
I processes having NOSWAP set. A MOSWAP set indicates that that process cannot be swapped out. If such a process exists, has greater than five percen~ of the user memory, and is in the COM
i mode less than some percentage of the time, such as 10% then rule , 22 triggers. As indicated in the message template for this rule in Appendix 1, swapping is suggested. Also printed out with the . 3 ~ 4~7 message template for rule 22 would be the user name, image name, time, number of free balance set slots, size of the free list, working set size and in-swap-rate.
If the result at decision point 970 is no, then decision point 972 is reached. At decision point 972 questions are asked whether the larger of 2 or 5% of the process have more than 5% of memory~ rf not, rule 24 triggers indicating excessive overhead caused by swapping. Also printed out in addition to the message template in for this rule would be the time, in-swap-rate, number 'of free balance set slots, and the number of memory pages avail-able to users.
¦ If the determination in node 972 i~ yes, indicating no ex-cessive swapping, then decision point ~74 is reached, which ques-~ tions whether there is any process with a working set size 1 greater than its corresponding quota. If there are such pro-I cesses, than rule 23 is triggered indicating that there has prob-l ably been too much borrowing. As the message template for this ! rule specifies BORROWLIM and GROWLIM should be increased. Also II prin~ed out with the template for this rule, would be the user name, working set quota, image name, working set size, time, num-ber of ree balance set slots, and in-swap-rate.
If the answer at decixion point 974 is no, then decision point 976 is reached. If, at decision polnt 976, the number of COMO processes is ~ound to be greater than 50 percent of all the processes (i.e. most of the processes are computable), then decision point 978 is reached.

I-f at decision point 97~ the number of COMO processes at bpri (the lowest priority) is found to be greater than or equal to the number of COMO processes minus 1 (i.e., the COMO processes are at~base priority), then decision point 980 is reached.
At decislon pint 980 the sum of the working set sizes for all the batch jobs are compared to some percentage of usable mem-ory, such as 30 percent. If the sum of the working set sizes is greater than that percentage, indicating that there are large batch jobs, then rule 25 triggers. The associated message tem-plate appears in Appendix 1 and, when printed or displayed, would be accompanied by the time, in-swap-rat~, number of free balance set slots, number of memory pages available to users, and number ! of computable processes.
! If the answer at decision point 980 is no~ then rule 26 I fires, suggesting that to reduce overhead due to swapping, the I amount of time between swaps, i.e., SWPRATE, should be increasedO
¦ The template for rule 26 is also in Appendix 1, and the time, in-i swap-rate, number of free balance set slots, number of memory pages available to users, and nwmber of computable processes ~ ¦I would be printed out with that template.
If the result of decision point 978 was that all the COMO
proce~ses were not at base priori~y, then dec;sion point 982 is , reached, at which the difference between the total number of page faults and the valid faults is compared to some threshold number, 2S such as 25. The valid faults indicate the number of page faults ~ 814'~'7 caused by swapping. 'rhe to~tal p~ge faults, include both hard and soft page faults. If the difference is greater than the thresh-old, rule 27 triggers. The message template for that rule indi-cates ~xcessive overhead caused by swapping, and makes certain recommendatiohs, such as adding memory or reducing the working set sizes. Also printed out with the template would be the time, the in-swap-rate, the number of free balance set slots, the num-ber of memory pages available ~o users, the number of computable processes and the total fault rate.
If ~he determination at decision point 982 is no, then rule , 28 trigqers indicating that the system is 5wapping, rather than i adding pages. The m~ssage template for this rule contains sug-i gestions for changes in ~SQUOTA, PFRATH, WSINC to alleviate this I problem. The data to accompany that template include time, in-¦ swap-rate, number of free balance set slots, number of memory pages available to users, and number of computable processes, as well as total ~ault rate and gvalid fault rate.
If, at decision point 976, it was found that most processes ! were not computable, then the decision point 984 would be 1 reached, which determines whether the SWPOUTPGCNT is too large, such as greater than lOO. If so, rule 29 i5 triggered indicating that excessive overhead caused by swapping has occurred. The ¦¦ message template for this rule suggests a reduction of Il SWPOUTPGCNT. The time, average working set size, maximum working set size, in-swap-rate, number of free balance set slots, the ~L~83L4~7 number- of memory pages available to users/ and number of comput-able processes would accompany this templat~, If the result of the determination at decision point 984 is that SWPOUTPGCNT was not too low, then rule 30 triggers. The message template for this rule is that too much memory is being wasted on idle processes, and changes in WSQUOTA, PFRATH, WSINC, and SET PROC/NOSWAP are suggested~ Accompanying that template would be the time, average working set size, maximum working set size, in-swap-rate, number of free balance set slots, number of memory pages available to users, and number of computable pro-cesses.
If, back in decision point 960, it was determined that there !
was not too ~uch swapping, than the scarce free memory must be ! investigated, which happens at decision point 9~6. At that deci-~
I sion point, the sum of the working sizes of certain processes are ¦ compared to some percentage of usable memory, such as 25 percent.
Those processes whose working set sizes yo into the sum include those with sufficiently small page fault rates and whose sizes are less than WSQUOTA. If this condition is met, then rule 31 j triggers since allocation of memory seems possible. The message template for rule 31 suggests a decrease in ~SQUOTA of ~hose large users. The template for rule 31 also shown in Appendix 1, would be printed or displayed along with the time, in-swap-rate, l size of the free list, total fault rate, user name, working set ; quota, working set size and image fault rate.

~43-'7 I-f the determination at decision point 986 were that alloca-tion of memory did not seem possible, then the free list size is compared to FREEGOAL at decision point 988. If the free list size is greater than or equal to FREEGOAL, then stop point 990 is reached. Otherwise, rule 32 is triggered indicating that a prob-lem may soon occur. This is explained in the message template for rule 32 (Appendix 1) which would accompany the time, in-swap-rate, size of the free list, slze of the modified list, and total fault rate.
The rules for the CPU are shown in Figs. 10A and 10B and in-~! cludes rules 33-39. ~he first question that is raised in the . decision node 1000 is whether the COM and COMO processes togetheri are greater than or equal to 5. If not, indicating that there ¦ are not too many computable processes, than stop point 1005 is ~ reached. If the answer at decision point 1000 is yes, than at i decision point 1010, the determination is made is whether the highest priority user of the CPU has a higher ba e priority than ¦
another process who is in the COM mode more than one-third of the¦
tim~ and also has a sufficiently hi~h base priority, such as greater than 4. If these conditions are met, then a conclusion is reached that there is a higher priority lockout and the sug-gestion is made to adjust the base priorities, shown in the tem-plate for rule 33 in Fig. 12. If not, then at decision point 1015, the PRO_L_COM for all processes having a base priority of I greater than or equal to some number, such as ~, are added ' ~ 4~'~
together. The PRO_L~COM is the number of times that a process is found computable. If that sum is greater than or equal to another threshold, such as 5, and if there is a single process which is using more than half (or any other appropriate percent-age) of the CPU, than rule 4 i5 triggered, which indicates that there is a "hog" process which needs to be examined, as the tem-plate for rule 34 in Appendix 12 indicates. If the result at decision point 1015 is no, tha~ in decision point 1020, the in-terrupt CPU time is compared to some other thresholdt such as 20.
If the interrupt CPU time is sufficiently high, indicating that there is too much CPU time on the interrupt ~tack, than seYeral I dif~erent actions ~re sugg~ted inclu~ reconfiguration using ! different communication devices, such as the DMF 32 or DMZ 32, l using different interface~, changing the queued I/O's (QIO) into I
! a different type of operations and placing them into a buffer, orj redesigning programs that use video terminals. The template for rule 35 is also in Appendix 12.
If there is not too much CPU time on the interrupt stack, . than the determination mad~ in the next decision point, decision 1 point 1025, is whether there is any CPU idle time. If so, than l stop point 1030 is reached. If not, than decision point 1035 is ¦ reached. rn decision point 1035, the kernel time of the CPU is compared a~ainst some threshold, such as 30. The kernel time in l the VMS code relates to system services and overhead tasks. If the determination is that there is too much CPU kernel time, then 4~'~
decision point 1040 is reached`at which the variable QUANTUM is compared to a threshold, sUCh as 15, the total number of page faults is tested against another threshold, such as 100, and the total number of hard faults is tested against yet another thresh-S old, such as 10. QUANTUM is the amount of time that a process gets to run before it is scheduled out. If that time is suff i-ciently high and the hard and soft page fault rates are below a threshold, than the problem is-occurring with the kernel mode time perhaps due to the excessive use of some system services.
'The template for rule 36, which triggers if the determination at decision node 1040 is yes, is shown in Appendix 2.
, r the determination at dec1slon point 1040 is negative, . than the QUANTUM variable is checked against that same threshold i , at decision point 1045 to see if it is less than th~ threshold.
;If it is not, than stop point 1050 is reached. If it is, than ,I rule 37 triggers and the suggestion is to increase the QUANTUM
.¦ time as indicated in the template for rule 37 shown in j Appendix 2.
¦ If at decision point 1035 kernel time was not found exces-¦ sive, decision point 1055 is reached which tests to see if the ¦ Exec mode time exceeds a threshold, such a~ 20. If it does, ¦ rule 38 triggers and if not, rule 39 triggers. The message tem-plate for both rules appear in Appendix 2.
! Figs. llA-llD show the rules for I/O. The first determina-. tion that needs to be made at decision point 1100 is whether the 4~
disk having the most operatio~s per second is greater than or equal to a set threshold Ot' whether any disk has operations for second greater than or eq~al to the threshold. If so, then deci-sion point 1110 is reached at which the time spent in the file system (i.e., the files necessary to do I/O, is compared to the percentage of CPU time. If the time spent in the file system is greater than the percentage of CPU time, then decision point 1115 is reached.
At decision point 1115; the file cache hit ratio is compared ,to ~ threshold, such as 70, and the missed I/O rate is also com-pared to a threshold, such as 5. The file cache hit ratio is the - ~ ~of-tit~rt~ n~1~~b ~e-cbe-t~rnn~s~-andt he missed 1~0 rate is the number of times per second when the processes went out to Iget information about a file and was not able to find it. If the 1.
file cache hit rate is not too low, and if the missed I/O rate is not too high, then rule 40 is triggered. The message template ~for rule 40, which appears in Appendix 3, suggests a reconfiguration of the file system to produce overhead.
l Otherwise, i the file cache hit rate is too low, or the missed I/O rate is too high, then the decision point 1120 is ¦ reached in which the file open rate is compared to a threshold such as 5 per second. If the file open rate is too high, then ;¦rule ~0 triggers. The message template for rule 40 suggests Iminimizing the opening/closing of files, as indicated in the tem-;plate for rule 41 in Appendix 3.

-~7-1 ~ 8 ~4~'~
If decision point 1120 results in a no decision, than deci-sion point 1125 is reached in which the file header cache misses are compared to the total percentage of cache misses. If the file he~der cache misses are less than some percent, such as 90 percent, then stop point 1130 is reached, Otherwise, rule 42 is triggered suggesting an increase in file system caches, as the message template for that rule in Appendix 3 indicates. .
If at decision point lllOr the CPU time in the ~ile was no~
too hiqh, than decision point 1135 is reached, at which the exec-lO. utive mode time of the CPU is compared to a threshold, such as 200 If the executive mode time of the CPU is above that thresh-S-h~--~L~-4~3 i~ 9~eF-e~- ~9~eSt~n~-th-At there are exces-sive I/O demands with high executive mode times. Certain ,reconfiguration is recommended, in the message template for 15 1rule ~3 shown in ~ppendix 3.
. If, however, the CPU executive time is less than the thresh-old, than decision point 1140 is reached and the paging and . swapping operations per second to the disk with the most opera-l tions per second is compared to some percentage, such as SO per-20l cent of the total operations per second to that disk.
If it is determined that the disk is not doing primarily paging and swapping, then decision point 1145 is reached. At . decision point 1145, the top DIRIO (direct I/O) users dirio's per .
, second are compared to the total system dirio's per second. Ir the top user's dirio's per second is more than some percentage, ~ .28~7 such as 50 percent of the total system dirio's per second, and the queue on the disk of the most operations per second is greater than 1 5 ~average length), then rule 44 is triggered, suggesting again a reconfiguration especially of disk I/O. For details, see the message template for rule 44 in Appendix 3. If, at decision point 1145 it is determined that one user is not doing most of the I/O to disk, or the disk does not have a queue, than stop point 1150 is reached.
If in decision point 1140, it i5 determined that the disk is primarily doing paging and swapping, decision point 1155 is reached at which a determination is made whether the paging oper- ¦
l~l~e~~f tt~-tbr~r v~t~ t~ -m~st-~F~rat-ions per second ex-ceed some percentage, such as 50 percent of the operations per ,¦second of that disk. If so, it is determined that the disk is 15 I¦doing primarily paging, and rule 45 is trigge~ed suggesting cer-. tain reconfigurations listed in the message template for rule 45. i If at decision point 1155, the busiest disk was not found to ¦
be doing primarily paging, then decision point 1160 is reached.
At decision point 1160, the swapping operations to the disk with the most operations per second is compared to 50 percent of those operations. If it is determined that the disk is not doing , primarily swapping, stop point 1165 is reached. If the disk is found to be doing primarily swapping, than rule 46 is triggered indicating that the swapping is creating an excessive load on the disk, and suggesting further investigation, in accordance with l~he message template for rule 46 in Appendix 3.

-4g-~.~ 8~'7 I~, back in decision point 1100, the I/O rate to a disk was found not greater than the clevice threshold, than decision point 1105 is reached, which asks two questions. The ~irst is whether any disk has a queue greater than or equal to 1.5 (average). The second question is whether the FCP turns/file opens ratio is greater than or equal to a certain ratio, such as 3. The FCP
turns stands for the file control parameters which refers to psuedo cache accesses, and indicates the number of times that the system goes to a file to get information. If both of these con-ditions are met, then rule 47 triggers, and the method of thisinvention determines that there might be a fragmentation problem th~ dis~s. T~e-re~omme~*dLiuns-t~-~rrect such a problem are listed in the message template for rule 47, which appears in ~ Appendix 3.
lS ! If the result of the determination in decision point 1105 is ¦ negative, than decision point 1170 is reached, and the following l! two determinations are made. The first is whether the buffered ,;I~O rate is greater than or equal to some threshold, such as 70, ,I and th~ second determination is whether the sum of COM and COMO
20 j exceeds some threshold, such as 5. If both conditions are not I met, ~hen stop point 1175 is reached.
If, however, both conditions are met, than at dec;sion point 1180, the total terminal I/O rate is compared to another thresh-! old, such as ~0. If the total terminal I/O rate is greater thanthe threshold, then decision point 11~5 is reached.

~ Z'7 At decision point 1185, the total CPU interrupt time is com-pared to another threshold, such as 20. If the total CPU inter-rupt time is greater than that threshold, then decision point 1190 is reached.
s At decision point 1190, the presence of a DMF 32 or DMz 32 is sensed. If they are present, then rule ~8 triggers. The decision is that the terminal I/O is burdening the CPU with in-terrupts, and appropriate suggestions for reconfigurations are made, in accordanee with the message template for rule 48.
10 ~ If not, then rule 49 triggers suggesting a different kind of ¦ reconfiguration. This sugqestion is explained in the message Itemplate for r~le ~9 in Appendix 3.
¦ If at decision point 1185, it was dete~mined that the CPU
l! time on the interrupt stack was not too high, then decision poin~
¦1195 is reached. At decision point 1195, the CPU time in the ,¦kernel mode is tested against that theshold, such as 30. If the ~¦~PU time in the kernel mode is not too high, than stop point 119 i is reached. Otherwise, rule 50 triggers suggesting a redesi~n to reduce the large number of QIO'S. The details of the recommenda-20 ¦ tion are given for the message template for rule 50.
If, in decision point 1180, the I/O rate to the terminals , was less than the determined threshold, then rule 51 triggers, indicatin~ that some other terminals may be consuming the CPU re-source, as indicated in the message template for rule 51 in Appendix 3.

, The other rules are channel rules, resource rules and clus-ter rules. The one channel rule, rule 52, determine~ whether all of the I/O's per second on some CI (communications interface) port is greater than a threshold, such as 2125000, or whether or s the I.'O's per second on a single UBA (universal bus address) is greater than another threshold, such as 1000000, or whether the sum of all the I/O's per second on a single MBA Imass bus address) is greater than another threshold such as 1700000. If so, rule 53 triggers and, after the rule triggers a sufficient number of times, the message in the template for rule 52 (Appen-dix 4) is printed or displayed.
, In th~ pre~sL-~d-~mbcdh~mqrt Or th,-s--~-nvention, there are 'also several resource rules which described below as rules 53-70.
~The message templates for these rules also appear in Appendix 5.
15 ! I ~ule 53 determin~s whether for a particular node the sum of ! DECNET arriving local packets per seoond plus the DECNET
~departing local packets per second plus the DECNET transit pack-~jets per second is greater than or equal to some threshold, such ¦as 100. If so, than the determination is made that too many ¦packets are being handled by a particular node and rule 53 trig-g~rs.
The next rule, rule 54, questions whether any image in a Iprocess is walting for the mailbox ~RSN$MAILBOX) in order to transfer messages more than a certain number o~ times, such as 2.
If so rule 54 triggers and the associated message template in . . -5~-~ 4~'7 Appendix S is printed out or displayed after sufficient occur-rences.
Rule 55 determines whether any process is waiting for the nonpag~d dynamic memory (RSN~NPDYNMEM). If so, rule 55 triggers.
The correspon~ing message template for rule 55 is also in Appen-dix 5.
For rule 56, a determination is made whether any process was waiting for the paging file re~ource (RSN$P~FILE) o If so, rule 56 triggers.
10 , Rule 57 triggers when any process was waiting for the paged ,I dynamic memory (RSN$PGDYNMEM)o ~,¦ Ru ie ~ trlggers if any process is-wa-i~islg---for the resuurce ,¦ RSN$LOCKID. That resource refers to a lock identification ~jdatabase which can be used for communication between different 15 l¦ V~X computer systems in a cluster. If a process was waiting, . then the lock identification database is full. Th~ message tem-plate for rule 58, which is also in Appendix 5, explains this conditionc l For rule 59, a determination is made whether any process was 20¦ waiting for the swap file resource ~RSN$SWPFIL~) because that swap file space was full. If so, then rule 59 triqgers and the associated message template from Appendix 5 is printed out or displayed after certain number of occurrences.
In rule 60, a de~ermination is made whether any process was waiting for the resource modified page writer because that was ~ 8~ 7 busy (RSN$MPWBUSY) The modi~f;ed page writer resource writes the modified pages back on the disk. If so, rule 60 triggers and, after sufficient triggerings of this rule, the message template in Appendix 5 is printed out or displayed suggesting reasons ~or this condition and actions to be taken.
For rule 61, the determination is first made whether a pro-cess is waiting for the ~SN~SCS resource That resource is a communications protocol. If so, and this condition has occurred more than a certain number of times for an image, such as 2, then rule 61 trig~ers.
Rule 62 triggers when any process is waiting for the re-ource-R~*$~5TRA~ Tb~ nes~urce~*~termines whether the clus-;;
ter is in transition due to the addition or subtraction of a ~node.
lS I Rule 63 tests whether the page faults from the VMS operating i ;jsystem exceeds some threshold, such as three faults per second.
! If so, than suggestions are made for changing the working set ¦size of the ~MS if this rule triggers a sufficient number of ~¦times. The me~sage template for rule 63 containing those sugges-20 j tions is in Appendix 5.
In rule 64 the number of SRPs (small request packets) in ! use is compared to the parameter SRPCOUNT plus some percentage, ! like 5%. If ~o, then rule 6~ triggers.
Rule 65 and 66 are similar to rule 64, but compare the IRPs (intermediate request packets~ and LRPs ~large request packets), 4'~'7 respe~tively, to IRPCOUNT plus some percentage or LRPCOUNT plus some percentage, respectively. Those percentages may also be 5%
~ ule 67 triggers when the number of non-paged pool bites in use exceeds NPAGEDYN plus some percentage like 5%. MPAGEDYN is the number of non-paged dynamic memory bits which are preallocated to non-paged dynamic memory.
Rule 68 is the resource rule for the hash table. The number of resources in use is compared to the parameter RESHASHTBL, ; which is the number of entries preallocated to the hash table at 10 I boot time, plus some percentage, like 5%. If so, rule 68 fires.
The message template for this rule is also in Appendix 5.
ku~-69-triggers if -t~ umbe-r o~ ~e~ ~n use exceeds the j LOCKIDTBL plus some percentage, l ike 5%, Locks are used to man- ¦
¦ age shared reCources, and the LOCKIDTBL is a table of the possi-15 i,l ble lock identifications. I
l Rule 70 is related to the number of batch jobs in a system.
i In particular, rule 73 tests whether the batch jobs use more than ! a certain percentaqe, such as 70 percent, of the CPU for all ofl the intervals under consideration, and whether there are fewer than a certain number, such as 5 per interval, interactive jobs on the average. If so, then it appears that the CPU is running mostly batch jobs, and rule 70 triggers. The message template for rule 70 is also in Appendix 5.
In the preferred embodiment of this invention, there are also some network rules, called cluster rules. The cluster 8~
rules,-which test certain metr~ics and certain parameters for the VAXcluster, include rules 71~77. The message templates for these cluster rules are in Appendix 6.
For example, rule 71 asks whether the average queue on any disk is greatèr than a certain number, such as 1.2, and also asks whether the operations per second for any disk also exceed a threshold for that disk. If so, rule 71 triyqers and the message templa~e for that rule is shown in Appendix 6.
~ ule 72 is somewhat related and asks whether the queue on any disk exceeds a threshold, such as 2, and whether the opera-tions per second for that disk are less than another thresholdl -- - -- --., , _ _ __, .. , .. .. , j . If so, the cause Eor the rong ~ueue may not be heavy use, but may ~
be caused by partial hardware failure or path contention. Thus a j . different solution to the problem of queue length is suggested in 15 il th~ message template for rule 72.
,j Rule 73 ;s related to the HSC. The question is whether the j¦I/O bites per second for any HSC exceeds a certain number~ such as 3750000, or whether the operations per second for any HSC ex-1 ceeds another threshold, such as 500. If so, than the particular HSCs are over threshold and rul~ 73 triggers. The message tem-plate for that rule is in Appendix 6.
l Rule 74 triggers if any deadlocks are found. Deadlocks are .j caused because applications are using the lock manager during communications and incorrectly handling their own locking activ-ity. If rule 74 triggers a sufficient number of times, the rule 74 message template in Appendix 6 will be printed or displayedO

~ 4~7 For rule 75, three questions are asked. The first is wheth-er there were no deadlocks. The second is whether the number of deadlock searches exceed a threshold rate, such as .l per second.
The last is whether the parameter DEADLOCK_WAIT was less than another threshold, such as 5. DEADLOCK_WAIT is the amount of time ~hich the system waits before detecting whether a deadlock condition is present. If all three conditions were met, then there have been deadlock searc~es but no deadlocks found, which may be caused hy too low of a setting for the parameter DEAD-jLOCK WAIT. As the message.template for rule 75 indicates, a sug-;
gestion is to increase the value of that parameter.
! RU1~1 6~ha-s-~he~--~am~ hre~ sts-as rule 75~ but instead ! tests whether DEADLOCK_WAIT is greater than the threshold in rule I
l75. If so, the problems may be in the applications not in the 1.
15 'I DEADLOCX WAIT parameter, as the message template for rule 76 in-dicates.
The remaining cluster rule, rule 77, determines whether the number of disk I/Os caused by file cache misses exceeds som~. i ! threshold. If so, then the determination is that there have been ;
too many I~Os caused by a low hit ratio on the system file caches. If this rule triggers a sufficient number of times, the message template for rule 77, which appears in Appendix 6, is displayed or printed. The message in that template includes sug-gestions to examine the file cache statistics and perhaps get a !larger file cache.

~ ~ ~14~

Many of the rules 1-77 in`clude common characteristics, First, the triggering of a rule involves determinations of rela-tionships between thresholds~ metrics and parameters. Second, the message-template for a rule may not be printed out if the rule has not been eriggered a sufficient number of times. This avoids false alarms since the system will not siynal the presence of a problem if a rule triggered because of an artifact. Finally, the thresholds should be adapted to fit different networks, different processors, and different operating syst2ms.
10 ~ D Report Generation ~i There are two basic types of reports which are generated in ¦ accordancë with this lnvention. On is the ~tuning analysis" or I performance evaluation report which comprises the message tem-;! plates for the rules as well as the evidence which was evaluated ' 15,¦ in triggering the rule. The second type of report generated is a I work load characterization of the system or network, which is ¦ made possible by the data collected during the data collection ! process and by the organization of that data.
, The tuning analysis or performance evaluation report for a 201 rule is is generated if the number of times that a rule has been ! triggered exceeds a predetermined threshold for that rule. That report ;ncludes a message template for that the rule along with ! certain of the metrics, parameters, and thresholds examined in ! trig9ering the rule.

L4'~'7 Fig, 12 shows a preferred method for recording rule trig-gerings as well as the evidence for that rule. In accordance with the preferred embodiment of this invention, each performance problem rule, is associated with an evidence table, such as is shown in Fiy. 12. Each evidence table has as its root the count of the number of triggerings. The branches lead to a number of nodes at different hierarchical levels each representing a dif-ferent type of datum. The message template parallels the lev~ls and describes the type of datum at each level as well as the form of its presentation in the analysis report. As Fig. 12 shows, ~he transferred data, i.e., the data to be recorded when a rule ~-~~~ - ~ gers, 1~-3-tor-ed--i~ the evidence table.
~! Although Fig. 12 shows the evidence table as the tree struc-~ ture, and there are of course other ways of implementing a data 15 j~lstorage structure in the processor performing the method of this jlinvention.
In accordance with a preferred embodiment of the method of this invention, each time a rule triggers, a counter in the root (leve 1) of the evidence table associated with that rule incre-20 1 ments by one count. In storing the transferred data into the ev-idence table, the first datum of each set of data corresponds to the second highest level (level 2). That datum is compared to the other values in that same level. If there is a match, then a counter indicating the numher of occurrences of that value of the ! datum is incremented. If not, then a new entry, with a counter 4'~'7 set to 1, is stored at the second highest level with a branch back to the root.
If there is a match at the second highest level, then at the third highest level (level 3), the corresponding datum is com-s pared with the entries at that level which connect to the matchedentry at the second highest level. Again, if a match is found, the corresponding counter is incremented. Otherwise, a new tree entry is created branching back to the matched level above.
This procedure continues until all the transferred data has been stored at all the level~. Storing transferred data this way conserves memory space.
! Whënever a new en~try i-s-c-reated~~7~t~r--i-n the first or sub-sequent levels, the lower levels are not searched since there are i Ino connected entries at lower levels. Instead, entries are cre-15 jated at all the lower levels with counters set to 1.
I The storage of data in accordance with this preferred '¦embodiment of the invention is demonstrated in detail in Figs. ', ~13A-13D. Those fig~res show steps in the creation of an evidence i ,table for the transferred data in Fig. 12.
For the first triggering of the rule, the data entries are Al, ~1 and Cl, as shown in Fig. 12. As shown in Fig. 13A, the i counter at the root of the tree will be set to 1, and Al will be stored at the second highest level, Bl will be stored at the third highest level, and Cl will be stored at the fourth highest level. The data is stored in this manner since there could have been no match.

~x~
W~hen the next data entry ~ccurs due to the rule triggering, the counter at the root increments again. A search along the second level for Al occurs and a match is found, so the associ-ated counter increments to 2, as shown in Fig. 138. A search for S a Bl entry associated with Al then occurs. Since such an entry is found, that counter also increments to 2. The next search for a C2 entry, associated with both Al and 81 does not occur, how-ever, so a new entry for C2 is created with a count of l, For the set of data accompanying the fourth rule triggering, lO ~ th~ count at ~he root increments to four, as shown in Fig. 13C.
I A match is found at the second level for Al, so the counter in-..j.~
I crements to 4, but no match is~~und ~or a~~2 entry connected to ¦ Al. Thus, a new entry in the third level for B2 is created with ¦ a counter at l. When a search for C2 on the third level takes 15 lplace, no match is found connected to ~2. Thus, an entry for C2 Il is created with a counter of one. This is because the prior C2 ! entries corresponding to an Al/Bl path and this third C2 entry corresponds to an Al/B2 path even though there was a prior entry for C2.
20 j Fig. 13D ~hows the entire evidence table for the transferred data in Fig. 12.
. I Storing the evidence for a rule in this matter saves memory i space, since the amount of duplicate entries is reduced. In addition, the tree structure allows ease of access for report generation.

4~
~ nce the analysiS of the metrics and parameters is complete, the rule count is compared against an occurrence threshold for that rule. If the rule count exceeds that threshold, than the message template for the rule is printed out, and the evidence stored in the corresponding evidence table is printed out. Pref-erably, the evidence template for a given rule describes a line of text, which contains all the data in the hierarchical order in which they have been stored in the table. There is one line of text for each triggering of the rule, and for each line of text, ,there is one value o~ the data at each level retrieved from the evidence table and introduced into that line of text.
;l In àccordance with the preferred embodiment of this inven-'tion, the tree is traversed sequentially, and on each traversal, ! ~he counter of each value of the data at each level is ¦decremented and the value retrieved. The values remain identical `, ~¦until the lowest level counter is decremented to zero. At that ¦¦time, the next branch of the next h;gher node i~ taken. Figs. 14 and 15 show two examples of printouts or displays for the message 1~ template and the associated evidence.
In addition to the output oE evidence and messages, perfor-mance reports for the different processors and for the network as a whole can also be outputted. Performance reports provide dis-plays of work load characteristics to aid a system engineer or I manager in determining whether changes implemented based on the recommendations in the message template have improved or degraded . .

'7 system performance. Such reports include both tabular summaries and histograms, examples of which are shown in Figs. 16-18, Performance reports are produced using the data collected and stored in the database for anal~sis. That data includes both S processor ~etrics and paramater and network metrics. The presen-tation of such data in a performance report is an important part of the inventive concept, although the specific implementation of tabular or histogram printing ~an use well-known display pro-gramming techniques~
In general, performance reports provide statistics charac-I terizing processor and net.work work load, as well as resource L~
usagc by the active images. For example, a system manager re- .
', viewing a performance report could examine, for each image, aver- !
¦ age working set size, total number of page faults, elapsed time, ¦ percentage of CPU time, disk I/O, and terminal I/O.
Fig. 16, for example, shows an exemplory report for the interactive images run at one node. From this report a system manager can determine which ima~es are worth tuning considering, Ilfor example, both resoure usage and frequency of usage.
20 1 Other reports are also possible, such as showing the charac-teristics of all the types of processors.
Fig. 17 shows a performance report of network data, in par-li ticular, access to particular disks by all the nodes. This type i,of report is only possible because, unlike conventional systems, the method of this invention allows collection of data for an 4'~7 entire network. The difficulty of such data co~lection, including synchronization and data organization has been solved by the inventors thus allowing not only network-wide analysis, but presentation of network-wide performance data.
Fig. 18 shows a histogram which can be produced from the data available from the data collection operation. Histograms provides a ~raphic presentation of the work load characteristics for operator evaluation and anAlysis and provide a chronological jview of the CPU, disk, and terminal I/O usage for each node, as lO ~well as node status information.
¦ The data in histograms shows how the system is being used ~---t¦during speci~ied time intervals. ~ f-e-F~ reporting periods alter the scale of the histograms as well a~ their time resolu- ¦
, tion~ Such histograms help the system manager or engineer li lS i double-check the conclusions presented in the analysis section I and the recommendations effected.
For example, by analyzing a histogram of residence times for ¦
an image (i.e., the time between image activation and termina-tion), a systems engineer can track images which consume a great deal of certain resource, such as database update or application build.
Fig. la provides an example of the usefulness of histograms.
In Fig. 18, the vertical axis represents percentage of CPU time used by the processor, and the horizontal axis represents the number of hours reflected by the histogram, in this case 24. The 6~-~ 4 ~7 histogram shows, for example, that at 12:00 noon, interrupts and DECnet jobs used 5% of the CPU times and interactive ~obs used 85% of the CPU time.
E. Conclusion An embodiment this invention has been using a particular VAXcluster arrangement containing VAX computers using a VMS oper-ating system. Of course the method has broader applicabilities to other processors or processor networks. In particular, the specific names of the metrics and parameters will have to be adapted for each processor or network and the thresholds for the ,rules may need to be changed, as may certain of the rules them- :
selves.
j Persons of ordinary skill will recognize that modifications ,iand variations may be made to this invention without departing 15 ijfrom the spirit or scope of the general inventive concept. This ~invention in its broader aspects is therefore not limited ~o the ~;specific details or representetive methods shown and de5cribed.

ll -~5-APPENDIX 1:` MEMORY RULES

Rule No. essaqe Templates 1. There are excessive page faults from the following processes. This might happen if the application program is designed badly or i~ it is a special type of program such as an ~I type program or CAD
program, etc~ If it is possible to redesign the program, please do so, If it is ~ossible to in-crease its WSQ~OTA, ~t might help. Please note below the user(s) who is ~are~ causing the problem and the number of occurrences.
2~ There are too many imaqe activations causing too many page faults. This may be caused by a poorly written command procedure which activates ~oo many _ 15 jl___ _ _ _ pro~rams. L_sted__elowO the users who caused ex-¦ cessive image activations.
l 3. There are too many page faults associated with too i many image activations in the system as a whole.
i This might happen if many application programs were designed badly or by attempting to run too i many command procedures frequently.
,¦ 4. Excessive hard faulting may have been caused by i too small of a pase cache. Hard faults are more ',1 expensive than soft faults and are more likely to l occur if the page cache is too small~ !
¦ You can increase page cache by increasing the val-: ues of MPW_LOLIMIT, MPW_HILIMIT~ MP~_THRESH, FR~EGOAL and FREELIM. FREELIM should be approxi-mately equal to ~ALSETCNT ~ 20 and GROWLIM should be equal to FREEGOAL-lo FREE~OAL should be approximately equ~l to the larger of 3 * FREELIM
or 2 percent of m~mory. MPW_LOLIMIT should be the smaller of 120 or 3 * BALSETNT. Use AUTOGEN to . cause automatic adjustment of related AWSA
parameters.
¦ 5. The high rate o~ page faulting might have been lower if more memory could be allocated to 8~
Rule ND. Messaq~_Templates processes' working sets. If the page cache.were smaller, it is possible that the page faulting . would have been reduced because the excess pages - could have become part of the processes' working sets.
You can decrease page cache by decreasing the val-ues of MPW_LOLIMIT, MPW_HILIMIT, MPW_THRESH, FREEGOAL and FREELIM. FREELIM should be approxi-mately equal to BAL~ETCNT ~ 20 and GROWLIM ShOU1d be eqUa1 tO FREEGOAL-1. FREEGOAL should be approximately equal to the larger of 3 * FREELIM
or 2 percent of memory. ~PW_LOLIMIT should be the - smaller of 120 or 3 * BALSETCNT. Use AUTOGEN to cause automatic adjustment of re1ated AWSA
parameters.
~. The total system wide page fault rate (hard or _ _s.oftl_was high with little availabie memory on the ! free lis~ while certain (moderately idle) pro-:, cesses maintained large working sets. If the swapper could reclaim more of the pages from the idle processes' working sets, the total pagefault j rate may fall.
Decrease the WSQUOTA for the following users. If ; detached processes are causing the problem, . decrease PQL_DWSQUOTA.
, 7. While excessive page faulting was occurring, there i were some users running images which seemed to . want more memory than their WSEXTENT~ allowed. If.
¦ the WSEXT~NTs for these users was larger, there ! may have been less page faulting.
Increase the WSEXTENT for the following users. If detached processes are causing the problem, in-crease PQL_DWSEXTENT.
~. Exc~ssive page faulting exists while there is ample free memory~ The swapper is unnecessarily trimmin~ working sets to prepare for an in-swap.
If BALSETCNT were high enough, proces~es would not be unnecessarily swapped out, thereby eliminating this condition.

~ 7 R~ No. Messaqe _e~Plates Increase BALSETNTo This will prevent the swapper from trimming processes when there is ample memory for them to grow. This will also reduce the page ~ fault rate.
9. Exc`essive page faulting occurred when free memory was almos~ exhausted. AWSA attempts to allocate available memory to processes who need it when the free list is larger than BORROWLIM, however ~ro-cesses cannot utilize these loans if GROWLIM is larger than both BORROWLIM and FREELIM, I f VMS
can provide more memory to processes who are page faulting, the total page fault rate may decrease.
Decrease GROWLIM below BORRO~LIM. This will allow processes to use loans granted by AWSA. Appropri-ate values can be achieved for these parameters by using AUTOGEN.
. The current values of these parameters along with ;l other relevant parameters are:
i PFRATH: %%%%, BORR0~LIM^ %%~%~, GROWLIM: %%%%%, PAGE CACHE: %%%%%%, FRE~LIM: %%%9~%, FREEGOAL:
%%%%%, WSMAX: %%%%%%, 10~ Excessive page faulting is occurring and VMS can-.j not allocate addikional memory to users working sets. The Automatic Working Set Adjustment (AWSA) . feature in VMS is used to alter the allocation of memory to users' working sets. By increasing . ~SINC, VMS can allocate available memory properly,. `.
, and may reduce the page f au1t rate .
¦ Set WSINC to 150, the default SYSGEN setting.
Because WSIN~ is zero (meaning AWSA is turned.
off), processes' working sets cannot grow.
11. Increase the size of the SWAPPING file. Processes l are not able to grow causing degraded performance~ :
1 By increasing the SWAPPING file, the processes will be allowed to grow and use available memory.
12. Excessive page faulting occurred while there was excess ~ree memory. Page faulting may have been -6~-~8~4~7 Rule No. Messaqe Templatqs lower if processes' working sets were able to grow fas~er. AWSA seems slow to respond to allo~w pro-cesses' working set to grow.
Decrease the value for PFRATH
The current value of PFRATH: ##
13~ Excessive page faulting exists while there is available free memory. AWSA response seems to ~e slow in providir.g processs' working sets with addi~ional memory. -The page fault rate might be reduced by improving AWSA's response, and allowing wor~ing sets to grow faster.
, Increase WSINC to 150, the default SYSGEN setting~
.' 14. Excessive page faulting exists while there is ~ ava~ la~e~ -ee-memory.__AWSA response seems to b~
. slow in providing processes' working sets with additional memory. The page fault rate might be ! reduced by improving AWSAs response, and allowing i working sets to grow faster.
! Decrease AWSTIME to 20, or allow AUTOGEN to reset AWSTI~. AWSTIM is the amount of CPU time between consecutive working set adjustments.
lSr Excessive page faulting occurred while there was l no f ree memory. Idle processes have laryer work- I
1 . ing sets than the heaviest page faulting pro-cesses. Automatic working set decrementing will allow VMS to recover idle pages from working sets to that they oan be used by processes which are page faulting thc most~ thus reducing the page fault rate~
Turn on Automatic working set decrementing by se~ting WSDEC to 35 and PFRATL to 10.
16. Excessive page faultinq in a scarce memory situa-tion has occurred. The swapper is forced to trim working sets often, creating the faulting and additional overhead. ~elow are several alterna-tives, but you should be careful in implementing changes due to the excessive load on the memory.

--6g--. . .

IL4~
Rule No. ~ aqe ~emplates Increase SWPOUTPGCNT to a value that is large enough for a typical process on the system to use . as its working set size. This will disable - second-level swapper trimming and possibly favor swapping. The swapper considers idle processes to be better candidates for memory reclamation than active processes, therefore potentially yielding a more favorable level of overhead.
The ideal value for LONGWAIT is the length of time that accurately d;stinguishes an idle or abandoned process from one tha~ is momentarily inactive.
Typically this value is in the range of 3 to 20 seconds~ You would increase LONGWAIT to force the swapper to give processes a longer time to remain idle before they become eligible for swapping or trimming.
_ _17.__Excessive_paging existed which may be due to ex- ;
~ ~~~~~ ~l ~ cessive automatic~working set decrementing. If 20;~ processes' workin~ sets are decreased too much or ' too oftenJ there will be a sudden increase in page ,I faulting. If the amount of pages taken away from I the working set is smaller, or i the value of ! PFRATL is lowered, the page fault rate may be low 25 , ered.
i Decrease WSDEC and/or decrease PFRATL.

18. Excessive page faulting exists due to heavy memory l demand. If the performance is unacceptable, you ; can attempt to reduce demand. One method would be.
to lower MAXPROCESSCNT. Adding memory may be your : best bet, but make sure that this is a recurring . situa~ion and get a second opinion.
1~. Excessive overhead caused by swapping exists.
BA~SETCNT is causing some of the problem because there are no free balance set slots, ! 200 ~xcessive overhead caused by swapping occurred.
Swapping might be reduced hy lowering the size of ! the page cache which would provide more memory to users.

~ 28~4~7 ule No. Messaqe Templates You can decrease page cache by decreasing the val-ues o~ MPW_LOLIMIT, MPW_HILIMIT, MPW_THRESH, - FREEGOAL and FREELIM. FREELIM should be approxi-- mately equal to BALSETCNT + 20 and GROWLIM should be equal to FREEGOAL-l. FREEGOAL should be approxima~ely equal to the larger of 3 * FREELIM
or 2 percent of memory. MPW_LOLIMIT should be the smaller of 1~0 or 3 * BALSETCNT, Use AUTOGEN to cause automatic adjustment of related AWSA
parameters.
21. Excessive overhead c~used by swapping occurred.
Swappin~ might be reduced by eliminating or con-trolling processe~ which consume excessive memory.
, The following user and image are consuming too : much memory. Try to make preventive adjustments such as: 1) Suspending the process until a non - ¦ pe k time of da~,_ ~ Reschedule the proces~ for a I non peak hour, 3) Limit ~hat user's WSQUOT~.
!¦ 22, Excessive overhead caused by swapping occurred.
Swapping of active processes is heavier because a large idle process cannot be swapped out. This . situation can occur when the command SET PRO-. CESS/NOS~AP is used.
1 23. Excessive overhead caused by swapping occurred.
A~SA can aggravate swapping if borrowin~ is too generous. Make sure GROWLIM is equal ~o l F~EEWGOAL - 1, and BOR~OWLIM is higher than ! GRO~LIM~
I 80rrowing is too ~enerous: increase ~ORRO~LIM and GROWBIM.
2~. Excessive overhead caused by swapping exists, There are too many concurrent processes fighting for the available memory. You might lower MAXPROCESSCNT (and 8ALSETCNT accordingly), reduce . demand, or add memory. Before adding memory, ver-ify that performance is unacceptable and get a second opinion from an exper~.

~ 7 Rule_ND. Messaqe Templates 25. Excessive overhead caused by swapping occurred.
Many COM and COMO processes are competing for mem-ory.
Reduce the number of concurrent large batch jobs.
Reschedule them for when there is less demand or consider lowering JOBLIM on batch ~ueues. Adding memory may help, but be sure and get a second opinion first.
26. Excessive overhead caused by swapping occurred.
Becau~e most process@s are computable and at base priority, you can reduce the swapping rate by increasing the amount of time between swaps.
Consider increasing SWPRATE~
I 27. Excessive overhead caused by swapping occurred.
I f the performance is frequently unacceptable, h~ an-~xp~-~t~ ~Luate-the possible need ~or more memory.
¦ Reduce demand or add memory~ Reduction of working , sets may help also.
'1 28. System swaps rather than pages. You might take the following steps to ease the condition:
1. Reduce W5QUOTAS
2. Increase PFRATH
2S I 3. Decrea~e ~SINC
29. Excessive overhead caused by swapping occurred.
There is heavy demand on memory, however, idle memory may be available, but allocated to users' working sets by the parameter SWPOUTPGCNT. This is the amount of pages VMS leaves in users' work-ing sets when trying to recover idle memory.
Reduce S~POUTPGCNT to 60. This will allow VMS to recover additional memory from idle users' working sets when the memory is needed~
I

4~7 Rule No. Messaqe TèmPlates 33. Excessive overhead caused by swapping occurred.
Most processes are not computable which suqgests .. that too much memory is be ing wasted on idle pro-- cesses, Pos~ible remedies ;nclude:
1. Reduce WSQUOTAS
2. Increase PFRATH
3. Might decrease WSINC
4. Check for processes that have ~5ET
; PROC/NOSWAP"
.l 31. Memory is scarce but NOT causing a performance . problem, however.if the demand for memory in- i ~ reases, a memor~_bottleneck may occur. In that j case, reallocate memory by decreasing t~e WSQUOTAs o users who are using more than 5 percent of mem-. ory.
32. Memory is scarce but NOT causing a bottleneck, I
I however, if your capacity plan predicts a growth ! in demand, a memory bottleneck may occur. Be pre-ll pared.

i .

4X~
- APPENDIX 2: CPU RULES

Rule No. Messaqe Templates . .
33, There is an apparent bottleneck at the CPU due to the large number of COMO processes. ~here exists higher priority process(es) which are causing lower priority COM and/or COMO processes to wait for the CPU which may be the cause of the problem.
This is considered a LOCKOUT condition.
Examine an~or revie~ process priorities. For an equitable distrib~tion of CPU time over the COM
. . processes, be sure they all have the same ~ASE
; PRIO~ITY.
j 34. There is an apparent bottleneck at the CPu due to the large number of COMO processes. There is also ti ~ hog_~rocess consuming at least 50 percent of the cpu time.
I Examine the hog process for faulty design, misman-¦ aged priorities, or other possible reasons.
¦ 35. There are many processes waiting for the CPu.
, Also the CPU time on the interrupt stack is high.
i Interrupt stack time is CPU time which cannot be used by processes.
Four actions are suggested:
l 1. If your t~rminal I/O is not handled by a 1 DMF32 or DMZ32- Th~ DMA feature of the DMx32 I is benefici31 if the applicaiton program writes 200 or more characters at-a time.
l When the application writes more than 10 ! characters at a time, the solo transfer of ~ ~he DMx32 is more efficient.
; When applications write out less ~han 10 characters at a time, there is no signifi-cant peformance improvement of the DMx32 over the DZll.

L___ Messaqe TemPlates 2. IE you are using DZll or DZ32 interface, you might consider reducing the baud rate to re-duce the frequency with which the DZll or ~ DZ32 interrupts for another character.
3. Design applications to collect the QIOs into large W~ITE operations that write as many characters as oossible up to MAXBUF.
4. Design applications for video terminals to update the affected portions of the screen rather than desi-~ning applications that re-write the whole screen.
36. Kernel mode time has exceeded thresho~d~ It is ~ possible that excessive use of some SYSTEM SER- :
VICES will cause high kernel mode. To alleviate the problem, do one or more of the following:
1. Schedule the work load so that there are ! fewer compute-bound processes running concur-~l rently.
.. ¦ 2. See if it is possible to redesign some appli- !
:! cations with improved algorithms to perform .¦ the same w~rk using fewer sysirEM SE~VICES.
,1 If the above steps do not solve the problem, il 1. Multiple CPUs (clusters) may solv,~.your prob-¦1 . lem if your work load consists of independent jobs and data structures.
2. Get a faster CPU if your work load is such . that completion of e3ch piece d,epends on the ' completion of previous one.
3a l 3. Add more memory if there is a memory limita~ ;
tion.
37. Kernel mode time is too high; increase QUANTUM.
! Increase QUANTUM to at least 20 ms. The current l value for QUANTUM is ##, When QUANTUM is in-creased, the rate of time-based context switching 14V~7 Rule No. Messaqe ~ tes will decrease, and therefore the percentage of the CPU used to support CPU scheduling and associated memory management will also decrease. When this overhead becomes less, performance will improve.
38. Exec mode time has exceeded threshold. It is pos-sible that your I0 or database management systems are being misused. Investigate applications using ~0 systems such as RMS or RD8, for poor design or unbalanced parame~ers.
39. There is a CPU bottleneck because there are many processes waiting for the CPU. If memory or ~0 problems exist simultaneously, try to solve those problems first. If the performance i~ still unacceptable when this occurs and the problem per-sists, reduce your CPU demand or add CPU power.
To alleviate the problem9 do one or more of the following:
1. Schedule the work load so that there are I fewer compute-bound processes running concur-rent ly .
1 2. See if it is possible to redesign some appli-cations with improved algorithms to perform l the same work with less processing.
1 3. Control the concurrent demand for terminal I0.
I If the above steps do not solve the problem, mul-¦ tiple CPUs (clusters) may solve your problem if l your work load consists of independent jobs and 1 data structures, or you may need a faster CPU if your work load is such that completion of each piece depends on the completion of a previous one.
a~ sure and get a second opinion before purchasing a nev CPU.

~ ;314~7 APPENDIX 3: I/O RULES

Rule No. Messaqe Templates 40~ Reconfigure to reduce file system overhead. There is `excessive use of file control primitives.
41. Minimize the OPENING and CLOSING of files and use appropriate file allocations. Use the command $SHO DEV/FILES to determine which users are heavy users of files.
~2. Increase file system caches.
43, There is excessive IO demand with high ~XEC mode , time. IO runtime systems such as RMS and RD~ may ,¦ be your bottleneck. If it is possible, try to im-i prove your IO system file caching or file design _ _~_ __ ___other~e_r~configure to reduce IO demand or add 1 IO capacity.
, 44. Check user programs for too much explicit QIOs~ I
i If possible, attempt to redesign these applica- !
tions to make better use of file caching. This l might be accomplished by using RMS ins~ead of , using explicit QIOs.
45. Paging is creating an excessive load on th~ noted disk. This may be a memory related problem, how-l ever, by adding another paging file to an existing j disk, the situation may improve.
If you already have multiple paging files on sepa-rate disks, you might be able to move the paging files to less heavily used disks and/or faster disks in the configuration.
A VAX 11/780 can usefully use two paging files, and a VAX 8600 can use three. The primary pagefile should be kept small especially on a shared system disk. Put the secondary page files on separate disks. Use only one swap file, but 3S ! remember not to put it on a shared system disk either.

4~7 Rule No. Messaqe Tèmplates 46. Swapping is creatin~ an excessive load on the noted disk. This may be a memory related problem, however, if the swapping file is on a shared sys-~ tem disk, the situation can be improved by moving it to a disk on which other systems are not paging or swapping.
47. Queues of IO requests are formed on disks which have a low operation count. The window turn rate indicates there may be a fragmentation problem on the noted disks.
Refresh disks regularly and try to keep as much as 40% of the space on the disks empty. This keeps fragmenta~ion from occurring quickly and reduces , ~he seek time resulting in better IO performance.
~8. Terminal IO is burdening the CPU with hardware in--- i ter~pts~_ Rçduce demand or add CPU capacity.
49. Terminal IO is burdening the CPU with hardware in-I terrupts. Improvement might be achieved by 2~ 1 replacing DZlls with a device capable of burst ; output, such as the DMF32 or DMZ3~. The DMA fea-ture would allow for a more efficient use of com-munication facilities by reducing the software demand.
;l 50. Users' explicit use of QIOs for terminal IO is i burdening the CPU. If possible, try to redesign ! applications to group the large number of QIOs ! into smaller numbers of QIOs that transfer more characters at a time. A second alternative would be to adjust the work loa~ to balance the demand.
If neither o~ these approaches ;s possible, you need to reduce the demand or increase the capacity of the CPU.
51. It is possible that other devices ~uch as communi-1 cations devices, line printers, graphic devices, i non-DIGITAL devices or instrumention, or devices i that emulate terminals are consumin~ the CPU re-i source.

4;~'7 - APPENDIX 4: CHANNEL RULES

Rule_No. Messaqe Templates 52. Channel(s) over threshold. There are excessive IO
rates on the noted IO channels. Be cautious of bottlenecks due to the overload on the channels.

. I .
li . ', .1 . ', i ., !

!l -79.-~a~

- APPENDIX 5: RESOURCE RULES

R~le No~ Messaq___el~e~
. .
53. There is excessive number of Decnet packets han-dled by this node. We estimate that more than 30%
of CPU power is spent if there are more than 100 packets per second.
S4. The following images were waiting because the Mailbox wa~ full.
55. The following images were waiting because Nonpaged dynamic memory was full.
56. The following images were waiting bec~use Page file was full. If (Avera~e program size) *
(MAXPRO~ESSCNT) ~ 1/2 * (Paging file size) then ', - inC~.P~c~ si~e of_pag~_fi.Le_~.o 2 * (Average program :, size) t (MAXPROCESSCNT) ,! 57. The following images were waiting because Paged ! dynamic memory was full.
j 58. The following images were waiting because Lock `~! identification database was full.
59. The following images were waiting because Swap ' file space was full. If tAverage working set , ~uota of running processes) * (MAXPROCESSCNT) is ! . greater than or equal to 3/~ * (Swapping file ; size) then increa~e ~ize of swapping file to 4~3 *
1 (Average WSQUOTA) * (MAXPROCESSCNT~.
60. The following ;ma~es were waiting because ~odified page writer was busy or emptying.
A process which faults a modified page out of its ! working set is placed into this wait if the modi-: fied page list contains more pages than the SYSBOOT parameter MPW_WAITLIMIT.
. Typically, the reasons for this wait state are:

; -80~

~3.4'~'~
Rule N~. Messane Temolates 1. MPW_WAITLIMIT is not equal to MPW_HILIMIT.
2. A compute bound real time job may be blocking Swapper process.
S If none of the above are causing the problems, in-crease MPW _ HILIMIT and/or decrease MPW LOLIMIT.
The idea here is to make the difference between these two parameters large enough so that modified pages do not have to be flushed out often. Also make sure that MPW_HILIMIT is equal to MP~_WAITLIMIT.
61. The following images were waiting because of RSN$_SCS.
The lock m~nager places a process in this wait j state when the lock manager must communicat2 with -r~ its coun~erpar~s on other VAXcluster nodes to . obtain information about a particular lock re-. source.
. If a process is delayed in this wait state, it may j be an indication of CI problems, an unstable clus-ter, or loss of quorum, Check the console for messages that might indicate cluster status and check the error logs and consoles for information i about CI problems.
! 62. The following imag2s were waiting because of . . RSN$_CLUSTRAN, A process which issues any lock requests on any node of a cluster in tra~si~ion (that is, while a node is being added or removed) is placed into this wait state while the cluster membership sta-bilizes~
63~ The system fault rate for VMS is over 2 faults per second for the following t;me periods. Perfor-. mance can be improved for the whole system if the VMS fault rate can be reduced.
Increase the working set size for VMS (SYSMWCNT) : to reduce the system fault rate.

-Bl-~81~7 Rule No. Messaqe Templates 64~ Unnecessary overhead occurred to build additional SRPs from non-paged dynamic memory~ If more SRPs were preallocated at boot time, there would have - been no additional overhead incurred~ nor would there have been any wasted dynamic memory.
Increase the SRPCOUNT parameter to a value slightly higher than the maximum amount o SRPs in use. Set the SRPCOUNTV parameter to four times the value for SRPCOUNT (this will not hurt perfor-mance if set too high).
65. Unnecessary overhead occurred to build additional IRPs from non-paged dynamic memory. If more I~Ps were preallocated at boot time, there would have been no additional overhead incurred, nor would , there have been any wasted dynamic memory.
_ _ _ _ In-crea.se the IRPCOUNT parameter to a value sli~htly higher than the maximum amount of IRPs in .! use~ Set the IRPCOUNTV parameter to four times ~O ~I the value for IRPCOUNT (this will not hurt perfor-I mance if set too high).
¦ 66. Unnecessary overhead occurred to build additional ,¦ LRPs from non-paged dynamic memory, If more LRPs !I were preallocated at boot time, there would have j been no additional overhead incurred, nor would , there have been any wasted dynamic memory.
I Increase the LRPCOUNT parameter to a value i slightly higher than the maximuun amount of LRPs in.
; use. Set the LRPCOUNTV parameter to four times : the value for LRPCOUNT (this will not hurt perfor-. mance if set too high).
67. Unnecessary overhead occurred to build additional . non-paged dynamic memory. If more bytes were l preallocated to non-paged dynamic memory at boot j time, there would have been no addi~ional overhead incurred.
Increase the NPAGEDYN parameter to a value slightly higher than the maximum number of , non-paged bytes in use. Set the NPAGEVIR
parameter to three times the value for NPAGEDYN, L42'7 Rule No. Messaqe Templates 68. Unnecessary overhead occurred to build additional resource hash table entries. If more entries were - preallocated to the RESHASHTBL at boot time, there ~ would have been no additional overhead incurred.
Increase the RESHASHTBL parameter to a value slightly higher than the maximum number of known resources.
6~. Unnecessary overhead occurred to build additional lock id table entries. If more entries were - preallocated to the LOCKIDT~h at boot time, there would have been no additional overhead incurr~d.
Increase the LOCID~BL parameter to a value slightly higher than the maximum number of locks ' in use.
- , 70. Thi~_CP~U appears to be running mostly batch jobs and a very small amount of int~ractiv~ work~ If ' ~his is the case, you may want to consider i increasing the value o QU~NTUM. This will allow j jobs to consume more CPU cycles per context , switch, thus lowering overall overhead. On a i VAX-11/7aO, QUANTUM can be set as high as 500 for a batch environment.
The aYerage number of interactive users was %%%, ~5 1 and batch jobs consumed %%% percent oE the CPU.
.~

,1 '.

--~3--- APPENDIX 6: CLUSTER RULES

~ule No. Messaqe TemPlates . .
71. Queues are forming on heavily used disks. Longer delàys will be experienced when longer queues form.
Check the disk(s~ for possible fragmentation prob-lems. If a disk is close to capacity (more than 70% full) with many WRITE operatiQns, then frag-mentation is likely to occur quickly. If this is the case, re~resh the disk using the BAC~ utility without the /PHYSICAL qualifier.
i If fragmentation is not the problem, assign all new work duriny the noted times to other disk vol-i umes (if possible3, and attempt to lower usage on - 15-~ the-noted disk volumes.
72. Queues exist on disks which have a low operation I count, This suggests a delay caused by path con-tention or partial hardware failure.
,I Suggested remedy: Use SPEAR to make sure device error~ are not generated on the noted volumes, ! andJor attempt to alleviate any c~ntention to th~
. disk by adding paths (channels or controllersl.
ll 731 HSC(s) over threshold.
! 74. The lock manager has detected deadlocks occurringO
1 ~eadlocks are caused by applications which use the , lock manager and incorrectly handle their locking ; activity. Deadlocks can cause cluster wide per-formance degradation because deadlock searches are initiated at high priority and are not necessarily restricted to any one node.
Try.to isolate the application(s) which cause the deadlocks and redesign the locking algorithms.
75. Deadlock searches are occurring with no deadlock ! finds. Deadlock searches may be caused by too low 3S ; of a setting for the parameter DEADLOCK_WAIT.

3L~4~'7 Rule No. Messaqe TèmPlates Deadlock searches can cause cluster wide perfor-mance degradation, Deadlock searches are initi-ated at high priority and are not necessarily re-stricted to any one node.
Consider increasing the value of DEADLOCK _ WAIT to the AUTOGEN default (not less than 10 seconds).
76. The lock manager has detected deadlock searches occurring with no deadlock finds. It is likely lo that applications are holding restrictive locks for too much time (there by triggering a deadlock search). Deadlock searches can cause cluster wide : performance degradation. Deadlock searches are ! initiated at high priority and are not necessarily : restricted to any one node.
Try to isolate the application(sl which cause the j dead-~oc~ e~-c-he-s-and.. redesign the locking ,l algorithms to hold restrictive locks for as shor~ i ;j a time as possible.
;¦ 77. There are too many disk IOs caused by a low hi~ , ratio on the system file caches. When a file !-cache lookup fails, a disk operation is incurr~d . to retrieve the required data. It may be neces- !
I sary to increase the size of one or more file caches to reduce the extra I/O overhead.
! Examine the file cache statistics to determine if . a larger cache is needed. If you decide to in-crease the size of a cache, use AUTOGEN and reboot the system to cause the new ~ lues to be used.

-85-.

Claims (7)

1. A method of operating a data processor system to evaluate the data processor system's performance, the data processor system having a plurality of work load characteristics indicative of the data processor system's performance and also having a plurality of system parameters defining allowable operating ranges for the data processor system, the system parameters being capable of adjustment to affect the data processor system's performance and the evaluation being performed relative to a set of rules specifying acceptable states of the work load characteristics, the method comprising the steps of:
measuring, at least once during each of a plurality of predetermined time periods constituting major intervals, a value of each of a plurality of metrics corresponding to said set of rules, each of said metrics representing a measurable quantity in said data processor system, and said major interval being initiated periodically by said data processor system;
storing, in a memory of said data processor system, the value measured for each of said metrics and the major interval during which that measurement was made;
comparing the value stored for each of said measured metrics with a corresponding threshold for that metric stored in said data processor system memory, said comparisons taking place for selected ones of said major intervals;
evaluating the results of each of said comparisons between said stored values and corresponding thresholds to determine whether a plurality of predefined specific relationships for each of the metrics represented by said stored values and the corresponding thresholds for those metrics have been met, each of said specified relationships relating to one or more of said set of rules;
triggering the ones of said rules for which all of the related specified relationships have been met;
recording in said data processor memory data representing each of the rules which have been triggered and the ones of the major intervals during which each of those rules was triggered;
counting the number of said major intervals during which each of said rules has triggered, to obtain a total number of triggering major intervals for each said set of rules;
and recommending, from the rules which have triggered, the adjustment of certain ones of said system parameters to improve the performance of said data processor system, said recommending step including the substeps of examining a list of suggested actions stored in said memory to find the ones of said suggested actions for which the corresponding rules are rules which have triggered, said list of suggested actions including an identification of the ones of said rules corresponding to each of said suggested actions;
ensuring that the total number of said triggering major intervals for each of said rules exceeds a corresponding predetermined threshold number prior to displaying the suggested actions found in said examining step; and displaying the ones of said suggested actions found in said examining step.

2. A method of operating a data processor system to evaluate the data processor system's performance, the data processor system including a plurality of inter connected elements and having a plurality of work load characteristics indicative of the data processor system's performance and also having a plurality of system parameters defining allowable operating ranges for the data processor system, the system parameters being capable of adjustment to affect the data processor system's performance and the evaluation being performed relative to a set of rules specifying acceptable states of the work load characteristics, the method comprising the steps of:
measuring, at least once during each of a plurality of predetermined time periods constituting major intervals, a value of each of a plurality of metrics corresponding to said set of rules, each of said metrics representing a measurable quantity in said data processor system, and said major interval being initiated periodically by said data processor system;
storing, in a memory of said data processor system, the value measured for each of said metrics and the major interval during which that measurement was made;
comparing the value stored for each of said measured metrics with a corresponding threshold for that metric stored in said data processor system memory, said comparisons taking place for selected ones of said major intervals;
evaluating the results of each of said comparisons between said stored values and corresponding thresholds to determine whether a plurality of predefined specific relationships for each of the metrics represented by said stored values and the corresponding thresholds for those metrics have been met, each of said specified relationships relating to one or more of said set of rules;
triggering the ones of said rules for which all of the related specified relationships have been met;
recording in said data processor memory data representing each of the rules which have been triggered and the ones of the major intervals during which each of those rules was triggered;
counting the number of said major intervals during which each of said rules has triggered, to obtain a total number of triggering major intervals for each said set of rules;
and recommending, from the rules which have triggered changes in a specified configuration of said plurality of interconnected elements to improve the performance of said data processor system, said recommending step including the substeps of examining a list of suggested actions stored in said memory to find the ones of said suggested actions for which the corresponding rules are rules which have triggered, said list of suggestion actions including an identification of the ones of said rules corresponding to each of said suggested actions:
ensuring that the total number of said triggering major intervals for each of said rules exceeds a corresponding predetermined threshold number prior to displaying the suggested actions found in said examining step; and displaying the ones of said suggested actions found in said examining step.

3. The method of Claim 1 or 2, further including the step of outputting, for each of said displayed suggested actions, preselected ones of said metrics, parameters, and thresholds used or evaluating the results of said comparisons to determine whether to trigger the ones of said rules corresponding to said displayed suggested actions.

4. The method of Claim 1 or 2 further including the steps of:
constructing a display by said data processor system of a work load characteristic of said data process system from said stored metric values; and outputting said display.

5. The method of Claim 1 or 2 wherein said measuring step includes the substeps of measuring several values for selected rapidly changing ones of said metrics more than once during each said major interval, and averaging said measured values for said selected rapidly changing one of said metrics to obtain an average value for each of said selected rapidly changing ones of said metrics; and wherein said storing step includes the substep of storing the average values for said selected rapidly changing ones of said metrics in said data processor system memory.

6. The method of Claim 1 or 2 wherein said recording step includes the substep of constructing in said data processor system memory a structured database for each of said rules, said data base including information about selected ones of the metrics corresponding to said rules and the number of times each said rule has triggered.

7. A method of Claim 1 or 2, wherein said comparing step includes the substep of implementing said rules as a decision tree structure having a plurality of decision points each calling for a comparison to be made during said comparing step, and wherein said evaluating step includes the step of traversing said decision tree structure to make the comparisons called for at said decision points.