WO2013152216A1

WO2013152216A1 - Balancing database workloads through migration

Info

Publication number: WO2013152216A1
Application number: PCT/US2013/035309
Authority: WO
Inventors: Yun Chi; Vahit Hacigumus
Original assignee: Nec Laboratories America, Inc.
Priority date: 2012-04-04
Filing date: 2013-04-04
Publication date: 2013-10-10
Also published as: JP2015513333A; JP5914699B2

Abstract

A method for balancing database workloads among a plurality of database servers includes when a new server is available, selecting a set of master and slave databases to be migrated to the new server; and migrating the selected databases to result in a balanced new workload among all servers while minimizing migration cost; and during operation, if workload unbalance is detected in real time from a workload change in a certain database, iteratively selecting one database at a time to migrate to a different server to balance the workload.

Description

BALANCING DATABASE WORKLOADS THROUGH MIGRATION

This application is a non-provisional application of US Provisional Application Serial No. 61620151 filed 4/4/2012, the content of which is incorporated by reference.

BACKGROUND

The present invention relates to database workload balancing.

For various economic and business reasons enterprises are increasingly centralizing their backend computer systems in purpose built data centers. Data centers typically house high concentrations and densities of such computer systems and additionally provide databases to support customer needs. Data center operators have to make the decision to purchase the server boxes up-front and then provision resources on ever changing workload. Further, multiple different workloads may share resources on the same physical box and provisioning the workload requires taking into account physical constraints such as capacity constraints associated with the physical resources. The recent move towards cloud computing for data intensive computing presents unique opportunities and challenges for data center operators.

One key challenge that data center operators face is the provisioning of resources in the data center for specific customer workloads. For example, with a new server added, each existing server "donates" a set of masters and slaves to be migrated to the server so that

• there is no master-slave co-located on the new server,

• master-to-slave ratio are about the same across all the servers,

• memory capacities and CPU capacities are as balanced as possible across all the servers,

• amount of data being moved is minimal.

Another problem is hotspot elimination. For a given configuration of servers, if there are overloaded servers (a server is overloaded if it is a memory hotspot or a CPU hotspot, or both), resolve the overloading servers by migrating out one master from each overloaded server. Federated databases have been used where relations are split manually among different servers and a unified view is exposed to the clients. However, such a solution involves manual configuration, special merging and wrapper algorithms, and does not support dynamic rebalancing.

Another solution is to add and remove slave databases based on workloads. However, because slave databases (replicas) only support read queries and all the write queries have to go to the master, such a solution can be ineffective when the workload contains many write queries. In addition, adding slaves takes very long time, and so this solution is not able to handle quick workload changes.

Other approaches have assumed that the workloads are different for a master DB and its slaves. Therefore, the workload is balanced through swapping the roles or master and slave. However, such method is only effective when the workloads at master and slaves are greatly different and the swapping is supported by the underlying DB system. In another method, the load balancing is achieved by (a) changing resource allocation among different virtual machines at the same server and (b) adding and removing servers. However, the method assumes each DB instance is wrapped within a single virtual machine and each server must contain all DBs (either all masters of all DBs, or one slave from each DB), thus limiting the applicability of the method.

SUMMARY

In one aspect, a method for balancing database workloads among a plurality of database servers includes when a new server is available, selecting a set of master and slave databases to be migrated to the new server; and migrating the selected databases to result in a balanced new workload among all servers while minimizing migration cost; and during operation, if workload unbalance is detected in real time from a workload change in a certain database, iteratively selecting one database at a time to migrate to a different server to balance the workload.

Advantages of the preferred embodiment may include one or more of the following. The system achieves workload balancing in databases by using database migration in real time. Such a solution has fast response time, involves minimum manual intervention, and can use different features (such as memory and CPU footprint) and metrics (such as workload variance, LI -norm of workload variation, master/slave ratio in a server, and so on) in the target function. The system achieves workload balancing among databases and so the overall system performance is optimized among all databases and all the servers. Faster operation is achieved. The system leverages techniques of live migration of memory-resident databases and so the operation is fast and the reaction to workload changes is quick. The system quickly finds a cost-effective database arrangement that minimizes the cost of running a workload. The system helps data center operators to provision for specific workloads while minimizing the total operating cost (TOC) of running the workload. Workload balancing can be achieve in real time through migration, and the ACID properties (atomicity, consistency, isolation, durability), which are required for most database systems, are guaranteed.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates an exemplary workload balancing achieved by migration database masters or slaves among servers

FIG. 2A illustrates a low variance implying balanced capacities, for N = 2.

FIG. 2B is an exemplary illustration of low variance implying balanced capacities, for N

= 3.

FIG. 3 shows an illustration that the optimal tenant may not exist. FIG. 4 is an illustration of the hotspot elimination problem.

FIG. 5 is an illustration of a memory hotspot given multiple resources, namely memory and CPU.

FIG. 6 shows an exemplary computer to perform database balancing.

DETAILED DESCRIPTION

FIG. 1 illustrates an exemplary workload balancing achieved by migration database masters or slaves among servers. SV1-SV4 represent four servers, M1-M3 represent three database masters, and each database master has two corresponding database slaves (replicas). Thus, SI 1 and S 12 are slaves of Ml . First, when a new server is just started to handle more workload, the system selects a set of master and slave databases to be migrated to the new server. The selected databases to be migrated will result in the most balanced new workload among all servers (including the new server) while use the minimum amount of migration cost. In the new server, the master-to-slave ratio is similar to those in existing servers (i.e., the new server is as typical as any existing server after the migration is completed). Second, during operation, if workload unbalance is detected in real time, e.g., due to workload change in a certain database, our solution will iteratively select one database at a time to migrate to a different server, in order to balance the workload.

In one embodiment, M databases are hosted on N servers. Each database consists a primary DB (master) and ore or more secondary DBs (slaves), which are hosted on different servers. That is, a master DB cannot be co-located with any of its slaves on the same server, and neither can two slaves belonging to the same master be co-loated on the same server.

Each database has footprints on several system resources, such as CPU usage and memory usage, which can vary over time. Each server has a maximum resource capacity beyond which the server is considered overloaded. The resources are assumed to be additive, namely the server resource usage is the sum of resources usages of all the databases hosted on the server. Such an assumption is commonly accepted for CPU and memory usages. Preferably, the system performs live migration of DBs, e.g., the real-time migration capability from systems such as the NEC TAM. Next, terminologies are detailed:

• The system consists of N servers, SV_V ..., SV_N .

• Server SV_t contains K_t tenants, {t) , ... ,t ^!' } . (Note: Different servers may have different numbers of tenants). • Each tenant

is either a master or a slave. The master of a tenant cannot be co- located with any of its slaves on the same TAM server. has a footprint of memory usage ή mem and a footprint of CPU usage

• The memory usage and CPU usage are additive in that for server SV_t , the total memory usage is

.

• The units of memory and CPU are interchangeable among servers (e.g., memory unit being Giga-Byte and CPU unit being Giga-cycles-per-second).

• Each server SV_i has a memory threshold SV memTh and a CPU threshold SV cpuTh that in normal conditions, should not go beyond. Once going beyond either memory or CPU threshold, we say the server becomes a hotspot (a memory hotspot or a CPU hotspot).

• Each server SV_t has a memory capacity SV_i.memCap (namely, extra memory capacity before SV_i becomes a memory hotspot) and a CPU capacity SV cpuCap . If

SV memCap < 0 or SV cpuCap < 0 , server SV_i becomes a hotspot.

For each server SV_i .

• SV memCap: The memory capacity of server SV_i .

• SV cpuCap : The CPU capacity of server SV_i .

• SV memTh : The memory threshold of server SV_t , beyond which SV_i becomes a memory hotspot.

• SV cpuTh : The CPU threshold of server SV_i , beyond which SV_i becomes a CPU hotspot.

• SV ^tenants : The set of tenants {t), ... , t ' } that belong to server . For each tenant ή such that ή e SV_;.tenants ,

• t- Iid : Tenant ID of ή .

• ή .isMaster : A Boolean flag to indicate if

is a master (or a slave).

• ή mem : The memory usage of t- .

• t!jcpu : The CPU usage of ή .

• ή .dataSize : The size of data of

For N numbers {x ...,x_N} , the me

are defined as

1 ^N

^X N ¹'

1 ^N

With the above notation, a small variance implies balanced servers. If {x₁, ...,x_N} represent the resource capacities (memory or CPU) among servers, then lower variance among {x_x, ...,x_N} implies balanced servers. In another word, If remains unchanged, then

is minimized when χ_λ = · · · = x_N . σ_χ ² is a good metric to measure the balance of resource capacity among servers.

FIG. 2A illustrates a low variance implying balanced capacities, for Ν = 2. With x₁ + x₂ = c , where c = 2μ_x is a constant, the system simply redistributes values among χ_λ and x₂. Since x₁ + x₂ = 2μ_χ , all the possible {x₁ , x₂ } pairs are located on the line shown in the figure.

In FIG. 1, the equi-lines with the same

values are shown (they turn out to be circles). As can be seen, when x₁ = x₂ = μ_χ , σ_χ ² achieves its minimal value 0; other than that, the more unbalanced the {x₁,x₂} are, the higher the variance. This also works for N > 2. FIG. 2B is an exemplary illustration of low variance implying balanced capacities, for N = 3.

If SV_t is a hotspot server with resource (for example, memory) capacity x_t and SV_j is a light-loaded server with resource capacity je . . Then if the system migrates a tenant with resource usage Δ from SV_i to SV_j , the reduction in variance for capacities among servers is

/ (Δ) = ~~~ (^xi ^{~ x}j + Δ) . If this variance reduction is treated as a function of Δ , then

dA N N from which the best tenant to migrate should have resource usage Δ^* = (x . - _; )/2 and as a result of the migration, the change of variance is - (x . -x_; )²/4.

The solution is to choose the largest resource capacity. The tenant resource usage are discrete values and it is not guaranteed that a tenant with resource usage of Δ^* can be found, as illustrated in FIG. 3, where the optimal Δ^* is not achievable, because the nearest tenant resource usages are A₁ and Δ₂ . FIG. 3 shows this illustration that the optimal tenant may not exist.

Next, the process of new server migration is discussed. With a new server added, each existing server ^"donates" a set of masters and slaves to be migrated to the server so that

• there is no master-slave co-located on the new server,

• master-to-slave ratio are about the same across all the servers,

• amount of data being moved is minimal.

Pseudo-code for one exemplary Method 1 to migrate DBs to the new server is as follows: fo u : .S ^'j , .. , , 5¾ _Ϋ tte Ijois it s rver *>K

out ut: lusmm I fo ml ¾fo <.>ni &SM! target :>.;^·· v r ? o : gmt to

*~ :fV ·?···· id , ch ige <^«« o;

t for j 1. to JV^" do

. i:

mid

14 id

i» return C, >;

A single resource X can be picked for illustration purpose, where X = {x₁,...,x_N} are the resource usages of the existing N servers {SV ...,SV_N} . Now, a new server SV_N+1 is brought into the system. After each existing server donates some masters and slaves to the new server, assume the new resource usages in the N+l servers are Y = {y ...,y_N,y_N+i\ . Then the new mean and variance in resource usage are

Ν

μ_Υ = Uy and is not affected by the exact configuration of tenants.

^^Y N + l ^x

Method 1 can be used and invoked a number of times. The problem is different in that SV_N+l was initially empty. So to fill up the empty SV_N+1 quickly, the greedy nature of Method 1 may favor migrating large tenants existing servers, and therefore may result in large-tenant-only on SV_N+1 and at the same time, leave large ^''holes" in the existing server capacities (i.e., some servers have large tenants migrated out while others have no tenants migrated out).

Based on the above consideration, Method 2 ignores the variance at the new server. That is, at each iteration, the method searches for a tenant t such that by migrating t into the new server SV_N+1 , the variance reduction among the existing N servers (with respect to μ_γ instead of μ_χ ) is maximized: t= argmax - [(y_t - t_i ) - μ_γ ] ).

server i,tenantk m-ust ; {SV_\ . ... :> , 5!¾ ij _: di r qM ¾«?¾r , mi rmN am

ut t ^' il d iftd si

, , , ,S¾- S½._;.* }^■

i (ί ί«¾: ¾_;ί, ^''· ^:;°:°4 Q

s for t 4- I to N do

■ί mid

■S for ί ^■■■ Ϊ t ,:Y do

" j in_:: i-f ) dl- ^ i<^ifr_: i < ^a^>: . >_v;il:^. _i:.^ id , ; ^'·: » mid

,.<);

«> ;.r |5V_S .... , >d _Λ· , S%., , ]^■:

As shown above, Method 2 can be used for new server migration. In Method 2, among {S^ SV_N} , there are \ M\ masters and | 5^* | slaves. Then the method selects | M | l{N + 1 ) masters and | 5^* | l{N + 1) slaves to migrate to the new server SV_N+1. To select | M \ l(N + 1) masters, Method 2 is called with parameters reqMaster = true and reqNum =| M | l(N + 1) ; to select I S | l{N + 1) slaves, Method 2 is called with parameters reqMaster = false and

reqNum =| S | l{N + 1) . Q is a priority queue, which sort the candidates by the variance reduction. In method 2:

• Lines 2—5 compute the average resource usages μ_ηεη and μ_ηεηι (with new server SV_N+1 included).

• Lines 6—9 pick the best candidate tenant for migration for each server. Note that this candidate can be decided by each server locally, without consulting other servers.

• Lines 11-14 move the best candidate among N servers to SV_N+1 .

• Lines 15—24 update the best candidate tenant of each server. If slaves are being migrated, the slave newly moved to SV_N+1 may invalidate the best candidate in other servers, and so re-computation is needed.

Next, the problem of hotspot elimination is discussed. For a given configuration of servers, if there are overloaded servers (a server SV_i is overloaded if it is a memory hotspot with

SV memCap < 0 or a CPU hotspot with SV cpuCap < 0 , or both), resolve the overloading servers by migrating out one master from each overloaded server. This problem is illustrated in FIG. 1.

There are two resources (memCap , cpuCap) , instead of a single resource X . This will affect two things: (1) how hotspot is defined and (2) how the balance among servers is measured. For (1), a new server SV_i a hotspot if either SV_i.memCap < 0 or SV cpuCap < 0 . In either case, the hotspot has to be eliminated. This is illustrated in FIG. 5 which is an illustration of a memory hotspot given multiple resources, namely memory and CPU.

For (2), a simple solution can be used by combining the variance of memory capacity and that of CPU capacity:

2 2 2

<J = C mem - <J memCap + C cpu - <J cpuCap , ' where c_mem and c are parameters for calibrating the different units between memory and CPU (GB vs. G-cycle-per-sec).

In addition, c_mem and c can be used to control the emphasis on the balance of memory and that of CPU. For example, maybe CPU usage is more bursty and its balance is more important than that of memory usage, because the memory usage is relatively stable. A special case is to set

1

c mem ⁼ 2

'memCap and

1

By doing this, the system essentially normalizes both _memCap and _cpuCap to 1 and treats the balance of memory and CPU as equally important. This consideration is reflected in lines 7, 8 and 1 1 of Method 3.

Because different tenants have different data sizes, the costs of migrating them are different. One solution is to revise the objective function as the ratio between the variance reduction of migrating t and the cost of migrating t . For example, if the migration cost for t is proportional to t 's data size, the new metric is

t.dataSize where <J²(t) is the variance reduction if t is migrated.

A master cannot co-located with any of its slaves. So a master at a hotspot server cannot be migrated to arbitrary servers. This has to be checked in the algorithm. That is, for a tenant t who is a candidate for migration, the system only considers target servers that do not have t 's slaves on them. The system considers candidate tenant and target server that eliminate the old hotspot and do not generate new hotspot. If such a solution does not exist, the algorithm returns null.

The input of the algorithm is {SV ..., SV_N} . Among the servers, there might be more than one hotspots and so the solution of the algorithm should address all the hotspots. This makes the problem much more difficult. We choose to use a local greedy heuristic where the hotspots are eliminated one-by-one. That is, (1) each server hotspot is addressed by assuming it is the only hotspot, (2) the system configuration is hypothetically updated after each hotspot is addressed, (3) then the next hotspot is addressed by assuming the hypothetically updated system configuration, and (4) the algorithm declares failure if it cannot reach a final configuration where all hotspots are eliminated.

Note that this is heuristics because (1) we have to follow an order for which the hotspots are addressed and (2) for each hotspot we find a solution in a greedy manner where a less greedy solution may make next hotspot easier to eliminate.

FIG. 4 shows an exemplary process for balancing database workloads among several database servers by using database migration (201). The method includes migrating a set of databases from a set of existing servers to a newly available server so as to balance the workloads among servers (202) and a method of migrating databases among existing database servers so as to balance the workloads among the servers (209). The process uses a metric to measure the goodness of a given database configuration (203). The process also selects a sequence of database masters and slaves to migrate to the new server (204). In 205, the metric considers various system resources such as CPU usage and memory usage. In 206, the metric combines different factors in a unified way through a weighting strategy and a vector norm (Euclidean norm or LI norm).

Turning now to 204, the method includes selecting a set of master databases and slave databases to be migrated to the newly available server. In 207, while making the selection, the method takes into consideration factors including (a) the expect cost reduction after the migration, (b) the master-to-slave ratio at each of the server after migration, and (c) the migration cost that is related to the size of each database. In 208, the method chooses an optimal order of migration so as to minimize the negative performance impact on the system during the process of migration.

The method includes migrating databases among existing database servers so as to balance the workloads among the servers (209). A metric can be used to measure the goodness of a given database configuration (210). In 212, the metric considers various system resources such as CPU usage and memory usage. In 213, the metric combines different factors in a unified way through a weighting strategy and a vector norm (Euclidean norm or LI norm).

In an alternative, in 211, the method includes iteratively selecting the next database to migrate and the corresponding target database server, in order to eliminate all the hot spots. In 214, the method efficiently selects the next database to be migrated and the corresponding target server to be migrated to. In 215, the method iteratively migrates one database at a time until all the hot spots are eliminated.

os f S ¾ ho teS St, YarUo ioo r d lK-t ki

ι t^■*···^■ r< >/.?/ . skr" :.\;·;

.,^'t¾ ^■■■■ r u

if > S( ^£ h i r

c d

u« >m- : di _{. ;} , ,. _r :d %, n ns !ddh hore eordd he hors d

om iu:: sr^'ate !^'{^·£_? M get) : ¾. or ers! s*t of t rurrds to mlgrsr nd ¾ <^:> cs>m;?¾so3idmg ta ge i;em; s f r

- #;

o j > ZO;

emr * d; FIG. 6 shows an exemplary computer to perform database balancing. The system may be implemented in hardware, firmware or software, or a combination of the three. Preferably the invention is implemented in a computer program executed on a programmable computer having a processor, a data storage system, volatile and non-volatile memory and/or storage elements, at least one input device and at least one output device.

By way of example, a block diagram of a computer to support the system is discussed next in FIG. 6. The computer preferably includes a processor, random access memory (RAM), a program memory (preferably a writable read-only memory (ROM) such as a flash ROM) and an input/output (I/O) controller coupled by a CPU bus. The computer may optionally include a hard drive controller which is coupled to a hard disk and CPU bus. Hard disk may be used for storing application programs, such as the present invention, and data. Alternatively, application programs may be stored in RAM or ROM. I/O controller is coupled by means of an I/O bus to an I/O interface. I/O interface receives and transmits data in analog or digital form over communication links such as a serial link, local area network, wireless link, and parallel link. Optionally, a display, a keyboard and a pointing device (mouse) may also be connected to I/O bus. Alternatively, separate connections (separate buses) may be used for I/O interface, display, keyboard and pointing device. Programmable processing system may be preprogrammed or it may be programmed (and reprogrammed) by downloading a program from another source (e.g., a floppy disk, CD-ROM, or another computer).

Each computer program is tangibly stored in a machine-readable storage media or device (e.g., program memory or magnetic disk) readable by a general or special purpose programmable computer, for configuring and controlling operation of a computer when the storage media or device is read by the computer to perform the procedures described herein. The inventive system may also be considered to be embodied in a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.

The system has been described herein in considerable detail in order to comply with the patent statutes and to provide those skilled in the art with the information needed to apply the novel principles and to construct and use such specialized components as are required. However, it is to be understood that the invention can be carried out by specifically different equipment and devices, and that various modifications, both as to the equipment details and operating procedures, can be accomplished without departing from the scope of the invention itself.

Claims

What is claimed is:

1. A method for balancing database workloads among a plurality of database servers,

comprising: when a new server is available, selecting a set of master and slave databases to be migrated to the new server and migrating the selected databases to result in a balanced new workload among all servers while minimizing migration cost; and during operation, if a workload unbalance is detected in real-time from a workload change, iteratively selecting one database at a time to migrate to a different server to balance the workload.

2. The method of claim 1, comprising applying a metric to measure a goodness of a given database configuration.

3. The method of claim 2, wherein the metric comprises system resources including processor (CPU) usage and memory usage.

4. The method of claim 2, wherein the metric comprises different factors in a unified way through a weighting strategy and a vector norm.

5. The method of claim 1, wherein the selecting a set of master databases and slave databases for migration further comprises considering one or more factors including (a) the expect cost reduction after the migration, (b) the master-to-slave ratio at each of the server after migration, (c) the migration cost that is related to the size of each database.

6. The method of claim 1, comprising choosing an optimal order of migration to minimize impact on the system during migration.

7. The method of claim 1, comprising iteratively selecting a next database to migrate and a corresponding target database server to eliminate a hot spot.

8. The method of claim 1, comprising iteratively migrating one database at a time until all hot spots are eliminated.

9. The method of claim 1 , comprising determining a metric for a given database configuration including a mean and a variance, wherein the variance is determined as: mem memCap cpu cpuCap ^ where c_mem and c_cpu are parameters for calibrating different units between memory and a CPU.

10. The method of claim 9, comprising setting

1 1. A system for balancing database workloads among a plurality of database servers, comprising: a processor; code executable by the processor when a new server is available, including instructions for selecting a set of master and slave databases to be migrated to the new server; and migrating the selected databases to result in a balanced new workload among all servers while minimizing migration cost; and code executable by the processor during operation, including instructions for detecting if workload unbalance is detected in real time and iteratively selecting one database at a time to migrate to a different server to balance the workload.

12. The system of claim 1 1 , comprising code for applying a metric to measure a goodness of a given database configuration.

13. The system of claim 12, wherein the metric comprises system resources including processor (CPU) usage and memory usage.

14. The system of claim 12, wherein the metric combines different factors in a unified way through a weighting strategy and a vector norm.

15. The system of claim 1 1 , wherein the code for selecting a set of master databases and slave databases for migration further comprises considering one or more factors including (a) the expect cost reduction after the migration, (b) the master-to-slave ratio at each of the server after migration, (c) the migration cost that is related to the size of each database.

16. The system of claim 11, comprising code for choosing an optimal order of migration to minimize impact on the system during migration.

17. The system of claim 11 , comprising code for iterative ly selecting a next database to migrate and a corresponding target database server to eliminate a hot spot.

18. The system of claim 11, comprising code for iteratively migrating one database at a time until all hot spots are eliminated.

19. The system of claim 11, comprising code for determining a metric for a given database configuration including a mean and a variance, wherein the variance is determined as:

2 2 2

<J = C mem - <J memCap + C cpu - <J cpuCap , ' where c_mem and c are parameters for calibrating different units between memory and a CPU.

20. The system of claim 19, comprising code for setting mem 2 and c cpu

memCap