WO2016078178A1 - Virtual cpu scheduling method - Google Patents

Abstract

A virtual CPU scheduling method, which belongs to the technical field of computer virtualization, and solves the problems that I/O processing cannot be responded in time, load characteristics cannot be satisfied, and a load balancing policy is too simple in an existing scheduling algorithm. The present invention comprises an initialization scheduling step, a virtual CPU running queue inserting step, a virtual CPU running step, a load balancing step, a credit value and system load updating step, a credit value redistribution step, a physical CPU redistribution step and a virtual machine type changing step. In the present invention, virtual machines are divided into three types, different types of loads are dynamically isolated, and are bound to two groups of physical CPUs with different types of loads for running, and different time slices are given to virtual CPUs running different types of loads, thereby improving the running efficiency and ensuring the I/O performance. In the present invention, a load balancing policy is redesigned, and besides ensuring the isolation of different types of loads, a policy with lowest migration overheads is selected to solve the problem that a load balancing policy is too simple in an existing scheduling algorithm.

Description

Virtual CPU scheduling method [Technical Field]

The invention belongs to the field of computer virtualization technology, and particularly relates to a virtual CPU scheduling method, which meets various application requirements by setting multiple types of virtual machines, and dynamically load-isolated and optimized load balancing strategies greatly enhance the entire virtual The performance of the platform.

【Background technique】

At present, computer performance has a qualitative improvement in terms of CPU, memory or network. Running a traditional computer architecture of an operating system on a server cannot make good use of these advantages, resulting in a very unsatisfactory utilization of the computer. According to a report by the international research firm International Data Corporation (IDC), in a typical data center of an IT enterprise, the average utilization rate of the basic physical architecture is only 10% to 15%, and most of the computer resources are not fully utilized. Secondly, As energy consumption and maintenance issues get more and more attention, virtualization technology has begun to receive attention from academia and industry.

In the computer world, virtualization usually means that computer software runs on a virtual platform rather than a real physical machine. Virtualization technology allows a computer platform to run multiple operating systems simultaneously, and its applications run in separate environments without affecting each other, while ensuring security and isolation while significantly improving computer productivity. Modern computers provide powerful support for virtualization technology, ensuring that multiple virtual machines running on a single physical machine can achieve sufficient performance and run separate operating system instances in each virtual machine, so Many vendors are starting to port their services and applications to the virtualization platform. After years of development, CPU virtualization and memory virtualization technologies have matured, and the development of I/O virtualization technology has lagged behind. Because of security considerations, virtual machine I/O operations require virtual machine monitors (virtual Intervention of the abstraction layer, while the competition between I/O resources and I/O response delay between virtual machines always restricts the improvement of I/O performance, affects the overall performance of the virtualization platform, and also limits virtualization. Further promotion and application of technology. Therefore, how to improve the I/O performance of virtual machines has become one of the research priorities of current virtualization technologies.

Currently, there are two main ways to improve the performance of virtual machine I/O. One way is to reduce the intervention of the virtual machine monitor. The intervention of the virtual machine monitor will affect the normal operation of the virtual machine, and context switching occurs, causing huge overhead; Another way is to optimize the scheduling of the virtual CPU so that the virtual machine can get a faster response. At present, the mainstream virtual CPU scheduling algorithm focuses on allowing virtual machines to share computer resources proportionally and fairly, and cannot guarantee real-time performance. In applications with high response time requirements, the response delay of events is closely related to the virtual CPU queue position. The delay is unstable and it is subject to extreme delays once in the latter position. In 2008, Diego Ongaro et al. In the article "Scheduling I/O in Virtual Machine Monitors" (see ACM SIGPLAN/SIGOPS Conference on Virtual Execution Environments, 2008), it is proposed to add a burst (BOOST) state in the credit value scheduling method of the virtualization platform Xen, when the virtual CPU When the blocking state is awakened by an I/O event, it enters a burst state, causing a virtual CPU scheduling. In this scheduling, the virtual CPU in a burst state preempts the physical CPU and processes the I/O event, but even The burst mechanism is only useful when the virtual CPU is woken up. Once other applications can continue to consume the time slice of the virtual CPU, the virtual CPU will not block, so there is still a response delay for most I/O tasks. The problem. In 2012, Cong Xu, Sahan Gamage et al. proposed different adoptions in the article "vSlicer: Latency-Aware Virtual Machine Scheduling via Differentiated-Frequency CPU Slicing" (see The 21st international symposium on High-Performance Parallel and Distributed Computing, 2012). The length of the CPU time slice to run virtual machines of different types of load, allocate a shorter CPU time slice for the virtual machine running the I/O type load, and allocate a longer CPU time slice for the virtual machine running the computing load, so that It can reduce the waiting time of virtual machines running I/O type load and increase its scheduled frequency to improve I/O performance. However, this method still has the following problems: (a) When the virtual CPU running the I/O type load and the virtual CPU running the computational load are in the same virtual CPU running queue, the former is greatly affected by the latter, once waiting for a calculation The virtual CPU of the type load will have a large delay. Therefore, this method can only reduce the delay to some extent. (b) A virtual machine has only one fixed CPU time slice, so only one type can be run. The load does not meet the need to efficiently run both compute and I/O loads in a single virtual machine. In 2013, Cong Xu, Sahan Gamage and others presented a further improvement in the article "vTurbo: Accelerating Virtual Machine I/O Processing Using Designated Turbo-Sliced Core" (see USENIX Annual Technical Conference, 2013), specifically from the physical CPU. A physical core is allocated for processing I/O events, and a virtual CPU is allocated to the virtual CPU based on the original virtual CPU. The virtual CPU is dedicated to running I/O load and is specially allocated. The physical core is scheduled. This solution eliminates the impact of computational load on I/O processing and further improves I/O performance, but there are still some shortcomings: (a) static physical CPU resource allocation, when the system's different types of load weights vary greatly The resource utilization is very low; (b) The virtual machine has two types of virtual CPUs, and when running a single type of load, the virtual CPU performance cannot be maximized. With the popularity of multi-core processors, the current load balancing strategy in the virtual CPU scheduling algorithm is also considered too simple. It only traverses the physical CPU in the system and looks for non-idle virtual CPUs in the queue for migration, without considering the overhead of migration. And further improve performance. In today's high-speed applications, bandwidth and I/O latency requirements are quite high, such as Ethernet Fibre Channel, which is widely used in IT enterprise data centers, for such pairs of response delays. For high-demand applications, existing virtual CPU scheduling algorithms are not applicable.

In order to clearly understand the present invention, the terms used in the present invention are explained below:

Physical Central Processing Unit (PCPU): A real processor unit in a computer system; a computer system can have multiple physical CPUs, and the computer system marks and manages each physical CPU through physical CPU bitmaps.

Physical CPU scheduling structure: In the scheduling algorithm, each physical CPU has a corresponding physical CPU scheduling structure, which stores a virtual CPU running queue on the physical CPU, and a secondary timer that is triggered periodically. If it is the first physical CPU scheduling structure initialized in the system, you need to mark the physical CPU as the primary physical CPU and set the primary timer for it.

Physical CPU Bitmap: An integer array in which each bit in the array represents a physical CPU of the computer. The specified physical CPU is marked by performing a set operation on a bit in the physical CPU bitmap.

I/O bitmap: A physical CPU bitmap type defined by the physical CPU bitmap user. Perform a set operation on a bit in the I/O bitmap to mark those dedicated to scheduling the running I/O load virtual CPU. The physical CPU establishes the affinity of the running I/O load virtual CPU and the 1-bit physical CPU in the I/O bitmap by the affinity setting function.

Calculate the bitmap: a physical CPU bitmap type defined by the physical CPU bitmap user. Perform a set operation on a bit in the calculation bitmap to mark those physical CPUs dedicated to scheduling the computational load-based virtual CPU. And the sex setting function to establish the affinity of running the computational load virtual CPU and computing the 1-bit physical CPU in the bitmap.

Affinity of the virtual CPU and the physical CPU: In the case where there are multiple physical CPUs in the virtualization platform, one or more virtual CPUs can be specified to be run through the system-defined affinity setting function. The physical CPU ensures that the one or more virtual CPUs will not be migrated to run on a physical CPU other than the one specified. For a virtual CPU with affinity set, the bit of the corresponding physical CPU in the affinity variable in the data structure is set to 1. For the virtual CPU without affinity, the affinity variable in the data structure is all Bits are all 1, indicating that they can be scheduled on all physical CPUs.

Virtual Machine: A complete computer system that runs through a software and has full hardware system functionality running in a fully isolated environment.

Virtual machine scheduling structure: In the scheduling algorithm, each virtual machine has a corresponding virtual machine scheduling structure, which stores information such as its virtual CPU number, virtual CPU linked list, and virtual machine credit value weight.

Virtual CPU linked list: Each virtual machine has a virtual CPU linked list that manages the virtual machine's Virtual CPU. The number of virtual CPUs in the virtual CPU linked list is the number of virtual CPUs allocated to the virtual machine, sorted according to the virtual CPU number.

Virtual CPU (VCPU): The CPU that the virtualization platform allocates to the virtual machine.

Idle Virtual CPU (IDLE VCPU): The virtual CPU used for placeholders, which does not complete the actual work, has the lowest priority of -3, and has an idle virtual CPU on the virtual CPU running queue of each physical CPU.

Virtual CPU scheduling structure: In the scheduling algorithm, each virtual CPU has a corresponding virtual CPU scheduling structure, which stores information such as credit value, priority, and virtual machine to which it belongs.

Virtual CPU running queue: Each physical CPU has a virtual CPU running queue, which consists of virtual CPUs running on the physical CPU in descending order of priority; virtual CPU priority values are from high to low -1 , -2, -3, respectively, indicating that they can be scheduled, not scheduled, and used only for placeholders;

Multi-core processor architecture: In today's multi-core processor architecture, a host often has multiple physical sockets, and each physical package encapsulates multiple physical cores. After using hyper-threading technology, each A physical core (core) includes multiple physical CPUs, and each physical core in the same physical package can share a third-level cache; a physical CPU in the same physical core can share a secondary cache; such a cache structure can greatly enhance the internal CPU. Improve the system performance by reading the hit rate of the data without having to look it up in memory or on the hard disk.

Virtualization platform: It is a software and hardware system that allows multiple virtual machines to run safely on a set of physical hardware. The companies that provide virtualization platforms are: Microsoft, Citrix, and Weirui.

Xen: It is a virtualization platform developed by the University of Cambridge Computer Laboratory. It is closely integrated with virtual machines and takes up less resources. It is known for its high performance and low resource consumption.

Virtual Machine Monitor (VMM): A virtualized abstraction layer running between the underlying physical server and the operating system that allows multiple operating systems and applications to share hardware, and a virtual machine monitor for each I/O device Establish a corresponding I/O request queue.

Credit: The core variable in the credit scheduling algorithm determines the priority of the virtual CPU and the length of the CPU time slice that can be obtained. The value range is -300 to 300. Only when the credit value of the virtual CPU is greater than 0, The virtual CPU can be scheduled. Each time the virtual CPU is scheduled, the credit value of the 10× time slice length is consumed. When the credit value of the virtual CPU is less than or equal to 0, the next credit value allocation is required.

I/O type load: I/O-based applications, the CPU load is very light, usually only runs for a short time, and requires more frequent response.

Computational load: CPU-based applications, usually with a CPU load of 100%, which is required to drop Low scheduling frequency and longest running time.

Hybrid load: In the operating environment, there are both I/O-based applications and CPU-based applications. CPU-based applications tend to be close to 100% of CPU resources. Applications that are dominated by /O operations are blocked when they are not used up, so in this environment, I/O-oriented applications are given higher priority and their scheduling frequency is increased.

Fibre Channel over Ethernet: Mapping Fibre Channel to Ethernet allows Fibre Channel information to be inserted into Ethernet packets, allowing Fibre Channel requests and data between servers and storage devices to be transported over Ethernet connections without specialization Fibre Channel structure.

[Summary of the Invention]

The invention provides a virtual CPU scheduling method for a multiprocessor virtualized environment, which solves the problem that the I/O processing cannot be timely responded in the virtualized environment, the load characteristics are not satisfied, and the load balancing strategy in the current mainstream scheduling algorithm is too Simple question.

A virtual CPU scheduling method provided by the present invention includes a scheduling initialization step, a step of inserting a virtual CPU running queue, a virtual CPU running step, a load balancing step, a step of updating a credit value and a system load, a step of reallocating a credit value, and reallocating a physical The CPU step and the step of changing the virtual machine type are characterized by:

(1) Scheduling initialization step: Create a physical CPU scheduling structure for each physical CPU: set a main timer and an adjustment timer for the main physical CPU, and set a secondary timer for other physical CPUs; create a virtual machine scheduling structure for each virtual machine : Initialize the virtual machine credit value weight, set the virtual machine type; initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU: initialize the load variable of each virtual CPU, the I/O task urgency, and allocate for each virtual CPU. Credit value; initialize I/O bitmap A _b and calculate bitmap A _c , the physical CPU in A _b is dedicated to scheduling virtual CPU running I / O type load, the physical CPU in A _c is dedicated to scheduling running computing load Virtual CPUs, which are complementary physical CPU bitmaps; initialize the global I/O load variable L _b and calculate the load variables L _c , L _b and L _c to record the I/O load and the calculated load in the system, respectively. The credit value resource consumed by the virtual CPU; setting the priority of each virtual CPU;

(2) Insert virtual CPU running queue step: If the current virtual CPU runs the computing load, it is directly inserted after all virtual CPUs of the same priority in the virtual CPU running queue; if the current virtual CPU runs the I/O type load, then press it The priority is sequentially inserted into the virtual CPU running queue. For virtual CPUs with the same priority, the I/O urgency is sorted from large to small.

(3) Step of running the virtual CPU: If the credit value of the virtual CPU of the current physical CPU is less than or equal to 0, trigger load balancing, and go to step (4); otherwise, run the virtual CPU of the team first, if the current physical CPU is the primary physical CPU. During the running process, the secondary timer reaches the time, the step (5), the main timer to the time, the step (6), the timer is adjusted to the time, the step (7) is performed, and the user issues a command to modify the virtual machine type. Step (8) is performed; the current physical CPU is a non-primary physical CPU. During the running process, the secondary timer expires, and the step (5) is performed. The user issues a command to modify the virtual machine type, and the step (8) is executed; the virtual CPU reaches the running. After the time, go to step (2) and re-insert the queue;

(4) Load balancing step: determine whether the corresponding bit of the current physical CPU in the I/O bitmap is 1, and the non-idle physical CPU with the I/O bitmap set to 1 bit as the target CPU group, otherwise the bitmap will be calculated. Set a non-idle physical CPU as the target CPU group; then select the optimal physical CPU from the target physical CPU group as the target physical CPU of the migration, and abandon the load balancing if the optimal physical CPU is not found;

After the target physical CPU is found, it is determined whether the current physical CPU has a corresponding bit in the I/O bitmap, and the virtual CPU that is the most urgent I/O task in the virtual CPU running queue of the target physical CPU is used as the target virtual CPU. Otherwise, the virtual CPU in the virtual CPU running queue of the target physical CPU that has the historical load closest to the difference between the target physical CPU and the current physical CPU historical load is taken as the target virtual CPU, and the target virtual CPU is deleted from the original queue and inserted into the current physical CPU. Virtual CPU queue leader; turn back to step (3);

(5) Update credit value and system load step: update the credit value and load variable of the current virtual CPU, and update the historical load variable of the relevant physical CPU, and determine the corresponding physical value of the physical CPU running the current virtual CPU in the I/O bitmap. Whether it is 1, is to update its I / O task pressing variables and global I / O load variables, go back to step (3); otherwise update the global computing load variables; go back to step (3);

(6) Re-allocating the credit value step: determining the priority according to the credit value of the current virtual CPU, recalculating the credit value quota allocated by each virtual CPU, and updating the credit value of all the virtual CPUs to determine whether the credit value of the updated virtual CPU is Change from non-positive to positive, then change its priority, and return to step (3); otherwise, the priority of the virtual CPU does not change, and directly returns to step (3);

(7) Reassign the physical CPU step: Calculate the number of physical CPUs running the I/O load and the number of physical CPUs running the calculated load according to the global I/O load variable and the calculated load variable; and modify according to the calculated physical CPU number Corresponding I/O bitmap and calculation bitmap, re-establish the affinity between virtual CPU and physical CPU; clear I/O load variable and calculation load variable to 0, and update all virtual CPU I/O tasks urgently Degree variable, return to step (3);

(8) to change the type of virtual machine steps of: reading a user command, according to which the virtual your crash ID, a virtual machine type and number of virtual Q CPU X _i, read the original virtual machine type S _i, according to the original virtual machine and a virtual machine type Q Type S _i , re-establish the affinity of each virtual CPU of the virtual machine with the I/O bitmap or calculate the bitmap, and modify the virtual machine type variable, and return to step (3).

The virtual CPU scheduling method is further characterized by:

(1) The scheduling initialization step includes the following sub-steps:

(1.1) Create a physical CPU scheduling structure for each physical CPU:

Determine whether to create a physical CPU scheduling structure for the first physical CPU, and then perform sub-step (1.2); otherwise, the rotor step (1.3);

(1.2) Mark the first physical CPU as the main physical CPU, and set the main timer and the adjustment timer for it. The main timer period T _main = 30 ~ 100ms, which is an integer multiple of 10ms, and is an integer of T _times . times, adjusting the timer period T _modulated = 1200 ~ 3000ms, T is an integral multiple of _{the main} rotor step (1.3);

(1.3) a secondary set timer period of the timer T _times as the secondary (1 ~ 3) × 10ms; the physical CPU load history variable is initialized to 0 H _p; H _p variable load history recorded on the physical CPU The sum of all the credits consumed by the running virtual CPU, the physical CPU number p = 1 ~ G, G is the number of physical CPUs in the system, G = 1 ~ 64;

(1.4) Create a virtual machine scheduling structure for each virtual machine:

In the virtual machine scheduling structure of each virtual machine, the corresponding virtual machine credit value weight W _{i is} initialized to 256, the corresponding virtual machine type S _{i is} set to 0, and its virtual CPU linked list is initialized to be empty;

The virtual machine number i=1～V, V is the number of virtual machines created by the user, V=1～100; S _i is 0 indicates a computing load, and S _i is 1 indicates an I/O type load, and S _i is 2 indicates a hybrid load, and the hybrid load is a load equivalent to a calculation type and an I/O type load;

(1.5) Initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU:

Add the virtual CPU to the virtual CPU linked list of the virtual machine to which it belongs;

Initialize the load variable H _{ij of} each virtual CPU to 0,

Initialize the I/O task urgency variable I _{ij of} each virtual CPU to 0, I _ij =0~1. The larger the I _ij value, the more urgent the I/O task is, I _ij =0, indicating no I/O load. ;

And assigning a credit value to each virtual CPU, wherein the credit value of the jth virtual CPU of the i-th virtual machine is C _ij = G × 10 × T _main × W _i / (W _T × M _i );

Among them, the total weight of the system

The number of virtual CPUs in the virtual machine is j=1～M _i , and M _i is the number of virtual CPUs of the i-th virtual machine, which is set by the user and is an integer smaller than G;

(1.6) Initialize the I/O bitmap A _b and the calculated bitmap A _c :

Establish I / O bit map bit map calculation and A _b A _c, A _b is the bits are initialized to all-0, A _c is initialized to all you, a system default initialization operation computational load;

(1.7) Upon initialization, all default virtual CPU computational load operation, establishing each of the virtual CPU calculates the bit map A _c is set a physical CPU affinity;

(1.8) setting the global I / O load variable L _b and calculating the load variable L _c , and initializing both L _b and L _c to 0;

(1.9) Initializing the priority P _{ij of} each virtual CPU to -1, indicating that it can be scheduled;

(2) Insert the virtual CPU run queue step, including the following substeps:

(2.1) determining whether the current virtual CPU has an affinity with the calculation bitmap A _c , and indicating that the virtual CPU is running a computational load, performing sub-step (2.2), otherwise indicating that the virtual CPU is running an I/O type load, the rotor Step (2.3),

(2.2) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPUs in the queue are compared, otherwise the sequence is compared backwards until the virtual CPU runs the queue tail; after the insertion is completed, go to step (3);

(2.3) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPU is compared in the queue, go to step (3); otherwise, perform sub-step (2.4);

(2.4) Compare whether the priority P _{ij of the} current virtual CPU is equal to the virtual CPU priority P _ij in the virtual CPU running queue, and then perform sub-step (2.5), otherwise the sequence is compared backwards until the virtual CPU runs the queue tail. After the insertion is completed, go to step (3);

(2.5) comparing the current value of the virtual CPU if I _ij I _ij is greater than the value of the virtual CPU operation queue virtual CPU, and a transfer step (3) before insertion then the current virtual CPU virtual CPU compares the run queue virtual CPU; Otherwise, the order is backward, comparing whether the priority P _{ij of the} current virtual CPU is greater than the virtual CPU priority P _ij for comparison, and then inserting the current virtual CPU into the virtual CPU running queue to compare the virtual CPUs and then transferring to step (3); Otherwise sub-step (2.4);

(3) The virtual CPU running steps include the following sub-steps:

(3.1) In the virtual CPU running queue of the current physical CPU, determine whether the team head virtual CPU credit value C _ij is greater than 0, then proceed to sub-step (3.2); otherwise, trigger the load balancing policy, and then go to step (4);

(3.2) Determine whether the target physical CPU group set by the first virtual CPU affinity of the virtual CPU running queue is the physical CPU represented by the 1 bit in the A _b bitmap, indicating that it is running the I/O type load and setting it The running time slice length T _s =1ms ~ 5ms; otherwise it indicates that it runs the computational load, set its running time slice length T _s = 30ms ~ 100ms;

(3.3) deleting the team head virtual CPU from the virtual CPU running queue;

(3.4) determining whether the current physical CPU is the primary physical CPU, if yes, performing sub-step (3.5); otherwise, performing sub-step (3.8);

(3.5) running the team head virtual CPU on the primary physical CPU, the secondary timer of the primary physical CPU, the primary timer, and the adjustment timer are simultaneously counted; and simultaneously monitoring whether the user issues a command to modify the virtual machine type, and then executing the step (8); otherwise sub-step (3.6);

(3.6) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.7); otherwise, removing the first virtual CPU from the primary physical CPU, and then moving to step (2);

(3.7) When the secondary timer has experienced the time T _times , the secondary timer is reset, go to step (5); when the master timer has experienced the time T _master , the master timer is reset, go to step (6); The timer has experienced the time T _adjustment , then adjust the timer reset, go to step (7); if no timer is triggered within T _s time, the virtual CPU continues to run for T _s , the team first virtual CPU from the main physical CPU Remove, go to step (2);

(3.8) Running the team head virtual CPU on the current physical CPU, and timing the secondary timer of the current physical CPU; and monitoring whether the user modifies the virtual machine type, if yes, performing step (8); otherwise, performing sub-step (3.9) ;

(3.9) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.10); otherwise, removing the first virtual CPU from the current physical CPU, and proceeding to step (2);

(3.10) when subjected to a secondary timer _times the time T, then the secondary timer is reset, go to step (5), if the timer does not trigger any time T _S, the virtual CPU operation time duration T _S, the team The first virtual CPU is removed from the current physical CPU, and the process proceeds to step (2);

(4) The load balancing step includes the following substeps:

(4.1) Select the physical CPU group for load balancing: determine whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, indicating that the virtual CPU running queue of the physical CPU is running I/O type. Load, the non-idle physical CPU set to 1 bit in the I/O bitmap A _b is used as the target physical CPU group for load balancing, and sub-step (4.2); otherwise, the physical CPU queue is running the computational load. Calculating a non-idle physical CPU set to 1 bit in the bitmap A _c as a target physical CPU group for load balancing, and performing sub-step (4.2);

(4.2) Selecting the optimal physical CPU from the target physical CPU group: traversing the target physical CPU group in the corresponding A _b or A _c , and determining whether to find the non-physical CPU on the same physical core The idle physical CPU is the target physical CPU of the migration, the rotor step (4.8), otherwise the rotor step (4.3);

(4.3) Re-traversing the target physical CPU group in the corresponding A _b or A _c to determine whether a non-idle physical CPU in the same physical package as the current physical CPU is found, and then sub-step (4.4), otherwise the rotor Step (4.5);

If (4.4) determines the non-physical CPU idle is greater than the current H _p H _p physical CPU, the physical CPU is the target physical CPU, the rotor step (4.8) of the migration, otherwise step rotor (4.5);

(4.5) in the corresponding target physical CPU group, sequentially determine backwards whether to find a non-idle physical CPU in the same physical package as the current physical CPU, then the rotor step (4.4), otherwise the rotor step (4.6);

(4.6) Re-traversing the target physical CPU group to determine whether to find a non-idle physical CPU on a different physical package than the current physical CPU. If yes, perform sub-step (4.7). Otherwise, mark the current physical CPU as idle and run idle virtual. CPU, rotor step (3.4);

(4.7) determining whether the H _{p of} the non-idle physical CPU is greater than twice the H _p of the current physical CPU, and the non-idle physical CPU is the target physical CPU of the migration, the rotor step (4.8); otherwise, the current physical CPU Marked as idle, running idle virtual CPU, rotor step (3.4);

(4.8) Select the appropriate virtual CPU to transfer on the target physical CPU:

Determine whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, and traverse the virtual CPU running queue of the target physical CPU, and use the virtual CPU with the largest I _ij value as the target virtual CPU of the migration, and the rotor step ( 4.9); H _p half of the difference between the target virtual CPU or physical CPU traverse the virtual CPU run queue, find H _ij closest to the target physical CPU H _p and trigger load balancing physical CPU as the migration target virtual CPU, rotor step (4.9);

(4.9) removing the migrated target virtual CPU from the virtual CPU running queue of the target physical CPU and inserting it into the head of the virtual CPU running queue of the current physical CPU, and the rotor step (3.2);

(5) The steps of updating the credit value and system load, including the following sub-steps:

(5.1) updating the credit value C _{ij of the} current virtual CPU: assigning the value of C _ij -10×T ₁ to C _ij ;

Update the load variable H _{ij of the} current virtual CPU: assign the value of H _ij +10×T ₁ to H _ij ;

Update the historical load variable H _p of the physical CPU running the current virtual CPU: assign the value of H _p +10×T ₁ to H _p ;

The credit value of a virtual CPU running for 1ms is 10, and T ₁ is the time elapsed since the last virtual CPU has been running since the last time, in units of ms;

(5.2) Determine whether the physical CPU running the current virtual CPU has a corresponding bit value in the I/O bitmap A _b , which means that the I/O task is run, and its I _{ij is} updated, and I _ij -1/(10 The value of ×T ₁ ) is assigned to I _ij , and sub-step (5.3) is performed; otherwise, sub-step (5.3) is directly performed;

(5.3) According to the current type of virtual CPU running load, the credit value consumed by the virtual CPU is added to the system load: it is determined whether the physical CPU running the current virtual CPU has a corresponding bit value in the bitmap A _b , and the global value is updated. The I/O load variable L _b , the value of L _b +10×T ₁ is assigned to L _b , the rotor step (3.4); otherwise the global calculated load variable L _{c is} updated, and the value of L _c +10×T ₁ is assigned L _c , rotor step (3.4);

(6) The step of reallocating the credit value includes the following substeps:

(6.1) determining whether the current virtual CPU credit value is greater than 0, then proceeding to sub-step (6.2); otherwise, setting the current virtual CPU priority P _ij to -2, indicating that the scheduling cannot be performed again, performing sub-step (6.2);

(6.2) Newly assign a credit value NC _{ij to} the virtual CPU of each virtual machine:

NC _ij = G × T _main × 10 × W _i / (W _T × M _i );

Where NC _ij is the newly obtained credit value of the jth virtual CPU of the i-th virtual machine;

(6.3) Update the credit value C _{ij of} each virtual CPU:

Assign the value of NC _ij +C _ij to C _ij ,

Judging whether C _ij >0 and P _ij value is -2, is to change the priority P _ij about the virtual CPU to -1, the rotor step (3.4), otherwise the priority P _ij about the virtual CPU does not change, the direct rotor Step (3.4);

(7) The process of reallocating the physical CPU includes the following substeps:

(7.1) Calculate the number of corresponding physical CPUs;

Number of physical CPUs running I/O type load N( _Ab ):

N( _Ab )=(L _b /(L _b +L _c ))×G;

The number of physical CPUs running the compute load, N( _Ac ):

N( _Ac )=G-N( _Ab );

(7.2) Determine whether the number of physical CPUs set to 1 bit in the I/O bitmap A _b is less than N( _Ab ), then proceed to sub-step (7.3), otherwise the rotor step (7.4);

(7.3) Transfer the number of physical CPUs from the physical CPU running the computational load to run the I/O type load: Let Y=N( _Ab )—I/O Bitmap A _b Set the number of physical CPUs in _Ab Set the first Y bits of the I/O bitmap A _b to 1 and clear the corresponding bit in the calculated bitmap A _c to the rotor step (7.6);

(7.4) Determine whether the number of physical CPUs set to 1 in the I/O bitmap A _b is equal to N( _Ab ), then the rotor step (7.7); otherwise, perform sub-step (7.5);

(7.5) Transfer the number of physical CPUs from the physical CPU running the I/O type load to run the computational load: Let Y=I/O bitmap A _b set the number of physical CPUs - N( _Ab ) , clear the first Y bits in the I/O bitmap A _b to 0, and calculate the corresponding position in the bitmap A _c , the rotor step (7.6);

(7.6) reset the virtual CPU affinity: reestablish all virtual CPU and I Run I / O type load / O bit map A _b is set in a physical affinity for the CPU;

Re-establish the operating computational load is calculated for all virtual CPU and bit map A _c is set in a physical affinity for the CPU;

(7.7) Clear the I/O load variable L _b and the calculated load variable L _{c to} 0, and update the I/O task urgency variable I _{ij of} all virtual CPUs: assign the value of 1/H _ij to I _ij , the rotor step ( 3.4);

(8) Steps to change the virtual machine type, including the following sub-steps:

(8.1) According to the modified virtual machine type command issued by the user, read the virtual machine ID, the virtual machine type Q, and the virtual CPU number X _i , and read the virtual machine type in the virtual machine scheduling structure according to the virtual machine ID. S _i ;

The modify virtual machine type command includes a virtual machine ID field, a virtual machine type field, and a virtual CPU number field, and the virtual machine ID field gives a virtual machine ID of a type to be modified; the virtual machine type field gives a virtual machine type Q to be modified. Q is 0 for the computational load, Q is 1 for the I/O load, and Q is 2 for the hybrid load; the virtual CPU number field is for the number of virtual CPUs used to run the I/O load in the hybrid virtual machine. This field is valid only when Q=2;

(8.2) judging whether Q=0, yes, proceeding to sub-step (8.3), otherwise performing sub-step (8.6);

(8.3) Determine whether S _i =0, if yes, keep the virtual machine type S _i unchanged, the rotor step (3.5); otherwise, judge whether S _i =1, then proceed to sub-step (8.4), otherwise perform sub-step (8.5 );

(8.4) Determining that S _i =1 means that the virtual machine running the I/O type load is converted into a virtual machine of the computational load, and the virtual CPUs of the virtual CPU and the I/O bitmap A _b are removed. affinity, to re-establish the virtual machine and the virtual CPU is calculated for all bit map a physical CPU affinity of a center _c, and the modified S _i is 0, the rotor step (3.5);

(8.5) Determining S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the computational load, and the virtual CPUs of each virtual CPU and the I/O bitmap A _b are deactivated. and resistance, the re-establishment of the virtual CPU which calculates the respective bit map a _c is set in a physical CPU affinity, and S _i 0 is modified, the rotor step (3.5);

(8.6) judging whether Q=1, if yes, proceeding to sub-step (8.7), otherwise performing sub-step (8.11);

(8.7) judging whether S _i =0, if yes, proceeding to step (8.8), otherwise performing sub-step (8.9);

(8.8) indicates that the virtual machine running the computational load is converted into a virtual machine running an I/O type load, and the affinity of each virtual CPU to the physical CPU that is set to one bit in the calculation bitmap A _c is released, and the All virtual CPUs of the virtual machine have affinity with the physical CPU set to 1 bit in the I/O bitmap A _b , and modify S _i to 1, the rotor step (3.5);

(8.9) judging whether S _i =1, yes, keeping the virtual machine type S _i unchanged, the rotor step (3.5); otherwise performing the sub-step (8.10);

(8.10) Judging that S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the I/O type load, and the virtual CPU of each of the virtual CPUs is calculated and the physical CPU of the bitmap A _c is set. And re-establish the affinity of each virtual CPU with the physical CPU set to 1 bit in the I/O bitmap A _i , and modify S _i to 1, the rotor step (3.5);

(8.11) Determine Q=2, judge whether S _i =0, if yes, perform sub-step (8.12), otherwise perform sub-step (8.13);

(8.12) Determining S _i =0, indicating that the virtual machine running the computational load is converted into a virtual machine running the hybrid load, and the first X _i virtual CPUs in the virtual machine are removed and the calculation bitmap A _c is set to 1 bit. The affinity of the physical CPU, re-establishing the affinity of the first X _i virtual CPUs in the virtual machine and the physical CPU set in the I/O bitmap A _b , and modifying the S _i to 2, the rotor step ( 3.5);

(8.13) Judging whether S _i =1, then proceeding to sub-step (8.14), otherwise performing sub-step (8.15);

(8.14) determines S _i = 1, represents the operation I / O type load into the virtual machine a virtual machine running the hybrid load is released after the virtual machine M _i -X _i virtual CPU and I / O bitmap The affinity of the physical CPU of the 1-bit is set in A _b , and the affinity of the virtual CPU in the virtual machine after the M _i -X _i virtual CPU is calculated and the physical CPU set in the bitmap A _c is set, and the S is _{i is} modified to 2, rotor step (3.5);

(8.15) Determine S _i = 2, keep the virtual machine type S _i unchanged, and rotate the step (3.5).

In the steps of the scheduling initialization step, the reallocation of the physical CPU step, and the change of the virtual machine type, when each virtual CPU is established with the I/O bitmap A _b or the affinity of the physical CPU set in the bitmap A _c is set, The affinity function can be set by the system definition;

When the virtual CPU is deactivated with the I/O bitmap A _b or the physical CPU of the bitmap A _c , the virtual CPU and I/ can be re-established by the system-defined setting affinity function. The O bitmap A _b or the affinity of the physical CPU set to 1 bit in the bitmap A _c is calculated.

In the existing scheduling algorithm, the focus is on the fair distribution of physical CPU resources to each virtual CPU, but this kind of fairness is only superficial, because different loads have different requirements for the operating environment, I/O type load. Usually, the execution time is short, but it is desirable to schedule low-latency and high-frequency. According to the current scheduling algorithm, the virtual CPU running the I/O type load is easy to cause high delay due to waiting in queue.

In order to meet different application requirements and ensure I/O performance, the present invention classifies virtual machines, delay-sensitive virtual machines exclusively run I/O type loads; non-delay-sensitive virtual machines exclusively run computational loads; A virtual machine that runs multiple types of workloads simultaneously. Allocating a long time slice for a virtual machine running a computational load reduces the overhead of environment switching; assigns a shorter time slice to a virtual machine running an I/O type load, and obtains a higher scheduling frequency; for a hybrid load The virtual machine needs to take into account the characteristics of the two types of load at the same time, and then perform internal virtual CPU partitioning, set a short time slice for the virtual CPU running the I/O type load, and set a normal time slice for the virtual CPU running the computing load. Let it run efficiently. In order to prevent the computational load from affecting the I/O performance, the present invention performs dynamic load isolation, which divides the physical CPU into two groups, a set of virtual CPUs that schedule I/O loads, and a set of virtual CPUs that schedule computational loads. The basis of the grouping is the weight of the total load of the system occupied by the I/O type load and the calculation type load. According to the adjustment timer, the physical CPU group is re-adjusted every certain period, and the system is guaranteed to be good under the premise of load isolation. flexibility. In order to solve the problem that the current virtual CPU adjustment cannot reflect the urgency of the I/O task, the present invention adds an urgency index of the I/O task to the virtual CPU, so that the urgent I/O event can be preferentially processed; The present invention also considers the problem that the existing virtual CPU scheduling algorithm load balancing strategy is too simple. According to the scheduling structure of the physical CPU, the virtual machine and the virtual CPU in the present invention, a load balancing strategy is redesigned, taking into account the hierarchical structure of the physical CPU. Selecting the optimal target physical CPU in the same physical core, the same physical package, and different physical packages in turn, in addition to ensuring the isolation of different types of workloads, the strategy of selecting the least migration overhead is also selected when the target virtual CPU of the migration is selected. Depending on the type of load, for a compute load, choose a scenario where the physical CPU run queue load is equivalent after migration; for I/O load, select an I/O task. The most urgent virtual CPU migration to ensure I/O performance.

[Description of the Drawings]

Figure 1 is a schematic flow chart of the present invention;

2 is a flow chart of the steps of inserting a virtual CPU running queue;

3 is a flow chart of a virtual CPU running step;

Figure 4 is a flow chart of the load balancing step;

Figure 5 is a flow chart showing the steps of reallocating a physical CPU;

Figure 6 is a block diagram showing the steps for changing the virtual machine type.

【detailed description】

The invention is further described below in conjunction with the drawings and specific embodiments.

As shown in FIG. 1, the present invention includes a scheduling initialization step, a virtual CPU running queue step, a virtual CPU running step, a load balancing step, an update credit value and system load step, a process of reallocating a credit value, a process of reallocating a physical CPU, and a change. Virtual machine type step;

(1) Scheduling initialization step: Create a physical CPU scheduling structure for each physical CPU: set a main timer and an adjustment timer for the main physical CPU, and set a secondary timer for other physical CPUs; create a virtual machine scheduling structure for each virtual machine : Initialize the virtual machine credit value weight, set the virtual machine type; initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU: initialize the load variable of each virtual CPU, the I/O task urgency, and allocate for each virtual CPU. Credit value; initialize I/O bitmap A _b and calculate bitmap A _c , the physical CPU in A _b is dedicated to scheduling virtual CPU running I / O type load, the physical CPU in A _c is dedicated to scheduling running computing load Virtual CPUs, which are complementary physical CPU bitmaps; initialize the global I/O load variable L _b and calculate the load variables L _c , L _b and L _c to record the I/O load and the calculated load in the system, respectively. The credit value resource consumed by the virtual CPU; setting the priority of each virtual CPU;

(2) Insert virtual CPU running queue step: If the current virtual CPU runs the computing load, it is directly inserted after all virtual CPUs of the same priority in the virtual CPU running queue; if the current virtual CPU runs the I/O type load, then press it The priority is sequentially inserted into the virtual CPU running queue. For virtual CPUs with the same priority, the I/O urgency is sorted from large to small.

(3) Step of running the virtual CPU: If the credit value of the virtual CPU of the current physical CPU is less than or equal to 0, trigger load balancing, and go to step (4); otherwise, run the virtual CPU of the team first, if the current physical CPU is the main Physical CPU, running, secondary timer to time, go to step (5), master timer to time, go to step (6), adjust timer to time, go to step (7), user issues modified virtual machine type If the command is executed, go to step (8); the current physical CPU is a non-primary physical CPU. During the running process, the secondary timer expires. Go to step (5). The user issues a command to modify the virtual machine type. Step (8); virtual CPU After the running time is reached, go to step (2) and re-insert the queue;

(4) Load balancing step: determine whether the corresponding bit of the current physical CPU in the I/O bitmap is 1, and the non-idle physical CPU with the I/O bitmap set to 1 bit as the target CPU group, otherwise the bitmap will be calculated. Set a non-idle physical CPU as the target CPU group; then select the optimal physical CPU from the target physical CPU group as the target physical CPU of the migration, and abandon the load balancing if the optimal physical CPU is not found;

After the target physical CPU is found, it is determined whether the current physical CPU has a corresponding bit in the I/O bitmap, and the virtual CPU that is the most urgent I/O task in the virtual CPU running queue of the target physical CPU is used as the target virtual CPU. Otherwise, the virtual CPU in the virtual CPU running queue of the target physical CPU that has the historical load closest to the difference between the target physical CPU and the current physical CPU historical load is taken as the target virtual CPU, and the target virtual CPU is deleted from the original queue and inserted into the current physical CPU. Virtual CPU queue leader; turn back to step (3);

(5) Update credit value and system load step: update the credit value and load variable of the current virtual CPU, and update the historical load variable of the relevant physical CPU, and determine the corresponding physical value of the physical CPU running the current virtual CPU in the I/O bitmap. Whether it is 1, is to update its I / O task pressing variables and global I / O load variables, go back to step (3); otherwise update the global computing load variables; go back to step (3);

(6) Re-allocating the credit value step: determining the priority according to the credit value of the current virtual CPU, recalculating the credit value quota allocated by each virtual CPU, and updating the credit value of all the virtual CPUs to determine whether the credit value of the updated virtual CPU is Change from non-positive to positive, then change its priority, and return to step (3); otherwise, the priority of the virtual CPU does not change, and directly returns to step (3);

(7) Reassign the physical CPU step: Calculate the number of physical CPUs running the I/O load and the number of physical CPUs running the calculated load according to the global I/O load variable and the calculated load variable; and modify according to the calculated physical CPU number Corresponding I/O bitmap and calculation bitmap, re-establish the affinity between virtual CPU and physical CPU; clear I/O load variable and calculation load variable to 0, and update all virtual CPU I/O tasks urgently Degree variable, return to step (3);

(8) to change the type of virtual machine steps of: reading a user command, according to which the virtual your crash ID, a virtual machine type and number of virtual Q CPU X _i, read the original virtual machine type S _i, according to the original virtual machine and a virtual machine type Q Type S _i , re-establish the affinity of each virtual CPU of the virtual machine with the I/O bitmap or calculate the bitmap, and modify the virtual machine type variable, and return to step (3).

The application background of the embodiment of the present invention is that the host has two physical packages, and each physical package has four physical cores, and hyperthreading is enabled. Each physical core contains two physical CPUs, the experimental platform operating system is CentOS-6.4, the virtual machine monitoring program is Xen-4.2.0, the kernel version is Linux-3.2.0, and the host runs 16 virtual machines, each virtual machine. 4 virtual CPUs, the virtual machine operating system is CentOS-6.4, and the kernel version is Linux-3.2.0.

Embodiments of the present invention include a scheduling initialization step, a virtual CPU running queue step, a virtual CPU running step, a load balancing step, an update credit value and system load step, a process of reallocating a credit value, reallocating a physical CPU step, and changing a virtual machine Type step:

(1) The scheduling initialization step includes the following sub-steps:

(1.1) Create a physical CPU scheduling structure for each physical CPU:

Determine whether to create a physical CPU scheduling structure for the first physical CPU, and then perform sub-step (1.2); otherwise, the rotor step (1.3);

(1.2) Mark the first physical CPU as the main physical CPU, and set the main timer and adjustment timer for it, the main timer period T _main = 30ms, the adjustment timer period T _tune = 1200ms, the rotor step (1.3) ;

(1.3) setting a secondary timer, the secondary timer period is 10 _times T; initializing the historical load variable H _{p of} the physical CPU to 0; p = 1 to 16;

(1.4) Create a virtual machine scheduling structure for each virtual machine:

In the virtual machine scheduling structure of each virtual machine, the corresponding virtual machine credit value weight W _{i is} initialized to 256, the corresponding virtual machine type S _{i is} set to 0, and its virtual CPU linked list is initialized to be empty;

Wherein, the virtual machine number i=1 to 16;

(1.5) Initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU:

Add the virtual CPU to the virtual CPU linked list of the virtual machine to which it belongs;

Initialize the load variable H _{ij of} each virtual CPU to 0,

Initialize the I/O task urgency variable I _{ij of} each virtual CPU to 0, indicating that there is no I/O load;

And assigning a credit value to each virtual CPU, wherein the credit value of the jth virtual CPU of the i-th virtual machine is C _ij = 16×10×30×256/(W _T ×M _i )=75; wherein, the total system Weight W _T =16384,

The number of virtual CPUs in the virtual machine is j=1～M _i , M _i =4, which is the number of virtual CPUs of the i-th virtual machine;

(1.6) Initialize the I/O bitmap A _b and the calculated bitmap A _c :

Establish I/O bitmap A _b and calculate bitmap A _c , which are complementary physical CPU bitmaps, initialize each bit in A _b to all 0s, and initialize each bit in A _c to all 1s, indicating that the system is initialized by default. Computational load;

(1.7) Upon initialization, all default virtual CPU computational load operation, establishing each of the virtual CPU calculates the bit map A _c is set a physical CPU affinity;

(1.8) setting the global I / O load variable L _b and calculating the load variable L _c , and initializing both L _b and L _c to 0;

(1.9) Initializing the priority P _{ij of} each virtual CPU to -1, indicating that it can be scheduled;

(2) As shown in Figure 2, insert the virtual CPU run queue step, including the following substeps:

(2.1) determining whether the current virtual CPU has an affinity with the calculation bitmap A _c , and indicating that the virtual CPU is running a computational load, performing sub-step (2.2), otherwise indicating that the virtual CPU is running an I/O type load, the rotor Step (2.3),

(2.2) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPUs in the queue are compared, otherwise the sequence is compared backwards until the virtual CPU runs the queue tail; after the insertion is completed, go to step (3);

(2.3) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPU is compared in the queue, go to step (3); otherwise, perform sub-step (2.4);

(2.4) Compare whether the priority P _{ij of the} current virtual CPU is equal to the virtual CPU priority P _ij in the virtual CPU running queue, and then perform sub-step (2.5), otherwise the sequence is compared backwards until the virtual CPU runs the queue tail. After the insertion is completed, go to step (3);

(2.5) comparing the current value of the virtual CPU if I _ij I _ij is greater than the value of the virtual CPU operation queue virtual CPU, and a transfer step (3) before insertion then the current virtual CPU virtual CPU compares the run queue virtual CPU; Otherwise, the order is backward, comparing whether the priority P _{ij of the} current virtual CPU is greater than the virtual CPU priority P _ij for comparison, and then inserting the current virtual CPU into the virtual CPU running queue to compare the virtual CPUs and then transferring to step (3); Otherwise sub-step (2.4);

(3) As shown in Figure 3, the virtual CPU runs the steps, including the following substeps:

(3.1) In the virtual CPU running queue of the current physical CPU, determine whether the team head virtual CPU credit value C _ij is greater than 0, then proceed to sub-step (3.2); otherwise, trigger the load balancing policy, and then go to step (4);

(3.2) Determine whether the target physical CPU group set by the first virtual CPU affinity of the virtual CPU running queue is the physical CPU represented by the 1 bit in the A _b bitmap, indicating that it is running the I/O type load and setting it The running time slice length T _s =1 ms; otherwise it indicates that it runs the computational load, and sets its running time slice length T _s =30 ms;

(3.3) deleting the team head virtual CPU from the virtual CPU running queue;

(3.4) determining whether the current physical CPU is the primary physical CPU, if yes, performing sub-step (3.5); otherwise, performing sub-step (3.8);

(3.5) running the team head virtual CPU on the primary physical CPU, the secondary timer of the primary physical CPU, the primary timer, and the adjustment timer are simultaneously counted; and simultaneously monitoring whether the user issues a command to modify the virtual machine type, and then executing the step (8); otherwise sub-step (3.6);

(3.6) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.7); otherwise, removing the first virtual CPU from the primary physical CPU, and then moving to step (2);

(3.7) When the secondary timer has experienced the time T _times , the secondary timer is reset, go to step (5); when the master timer has experienced the time T _master , the master timer is reset, go to step (6); The timer has experienced the time T _adjustment , then adjust the timer reset, go to step (7); if no timer is triggered within T _s time, the virtual CPU continues to run for T _s , the team first virtual CPU from the main physical CPU Remove, go to step (2);

(3.8) Running the team head virtual CPU on the current physical CPU, and timing the secondary timer of the current physical CPU; and monitoring whether the user modifies the virtual machine type, if yes, performing step (8); otherwise, performing sub-step (3.9) ;

(3.9) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.10); otherwise, removing the first virtual CPU from the current physical CPU, and proceeding to step (2);

(3.10) when subjected to a secondary timer _times the time T, then the secondary timer is reset, go to step (5), if the timer does not trigger any time T _S, the virtual CPU operation time duration T _S, the team The first virtual CPU is removed from the current physical CPU, and the process proceeds to step (2);

(4) As shown in FIG. 4, the load balancing step includes the following sub-steps:

(4.1) Selecting the physical CPU group for load balancing: determining whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, and then the non-idle physical CPU in the I/O bitmap A _b is made available. The target physical CPU group of the load balancing is performed in sub-step (4.2); otherwise, the non-idle physical CPU in the bitmap A _c is calculated as the target physical CPU group capable of load balancing, and sub-step (4.2) is performed;

(4.2) Selecting the optimal physical CPU from the target physical CPU group: traversing the target physical CPU group in the corresponding A _b or A _c , and determining whether to find the non-physical CPU on the same physical core The idle physical CPU is the target physical CPU of the migration, the rotor step (4.8), otherwise the rotor step (4.3);

(4.3) Re-traversing the target physical CPU group in the corresponding A _b or A _c to determine whether a non-idle physical CPU in the same physical package as the current physical CPU is found, and then sub-step (4.4), otherwise the rotor Step (4.5);

If (4.4) determines the non-physical CPU idle is greater than the current H _p H _p physical CPU, the physical CPU is the target physical CPU, the rotor step (4.8) of the migration, otherwise step rotor (4.5);

(4.5) in the corresponding target physical CPU group, sequentially determine backwards whether to find a non-idle physical CPU in the same physical package as the current physical CPU, then the rotor step (4.4), otherwise the rotor step (4.6);

(4.6) Re-traversing the target physical CPU group to determine whether to find a non-idle physical CPU on a different physical package than the current physical CPU. If yes, perform sub-step (4.7). Otherwise, mark the current physical CPU as idle and run idle virtual. CPU, rotor step (3.4);

(4.7) determining whether the H _{p of} the non-idle physical CPU is greater than twice the H _p of the current physical CPU, and the non-idle physical CPU is the target physical CPU of the migration, the rotor step (4.8); otherwise, the current physical CPU Marked as idle, running idle virtual CPU, rotor step (3.4);

(4.8) Select the appropriate virtual CPU to transfer on the target physical CPU:

Determine whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, and traverse the virtual CPU running queue of the target physical CPU, and use the virtual CPU with the largest I _ij value as the target virtual CPU of the migration, and the rotor step ( 4.9); H _p half of the difference between the target virtual CPU or physical CPU traverse the virtual CPU run queue, find H _ij closest to the target physical CPU H _p and trigger load balancing physical CPU as the migration target virtual CPU, rotor step (4.9);

(4.9) removing the migrated target virtual CPU from the virtual CPU running queue of the target physical CPU and inserting it into the head of the virtual CPU running queue of the current physical CPU, and the rotor step (3.2);

(5) The steps of updating the credit value and system load, including the following sub-steps:

(5.1) updating the credit value C _{ij of the} current virtual CPU: assigning the value of C _ij -10×T ₁ to C _ij ;

Update the load variable H _{ij of the} current virtual CPU: assign the value of H _ij +10×T ₁ to H _ij ;

Update the historical load variable H _p of the physical CPU running the current virtual CPU: assign the value of H _p +10×T ₁ to H _p ;

(5.2) Determine whether the physical CPU running the current virtual CPU has a corresponding bit value in the I/O bitmap A _b , and if so, update its I _ij and assign the value of I _ij -1/(10×T ₁ ) to I. _Ij , sub-step (5.3); otherwise directly proceed to sub-step (5.3);

(5.3) According to the current type of virtual CPU running load, the credit value consumed by the virtual CPU is added to the system load: it is determined whether the physical CPU running the current virtual CPU has a corresponding bit value in the bitmap A _b , and the global value is updated. The I/O load variable L _b , the value of L _b +10×T ₁ is assigned to L _b , the rotor step (3.4); otherwise the global calculated load variable L _{c is} updated, and the value of L _c +10×T ₁ is assigned L _c , rotor step (3.4);

(6) The step of reallocating the credit value includes the following substeps:

(6.1) determining whether the current virtual CPU credit value is greater than 0, then proceeding to sub-step (6.2); otherwise, setting the current virtual CPU priority P _ij to -2, indicating that the scheduling cannot be performed again, performing sub-step (6.2);

(6.2) newly assign a credit value NC _ij for each virtual machine's virtual CPU, NC _ij = 75;

Where NC _ij is the newly obtained credit value of the jth virtual CPU of the i-th virtual machine;

(6.3) Update the credit value C _{ij of} each virtual CPU:

Assign the value of NC _ij +C _ij to C _ij ,

Judging whether C _ij >0 and P _ij value is -2, is to change the priority P _ij about the virtual CPU to -1, the rotor step (3.4), otherwise the priority P _ij about the virtual CPU does not change, the direct rotor Step (3.4);

(7) As shown in FIG. 5, the process of reallocating the physical CPU includes the following substeps:

(7.1) Calculate the number of corresponding physical CPUs;

Number of physical CPUs running I/O type load N( _Ab ):

N( _Ab )=(L _b /(L _b +L _c ))×16;

The number of physical CPUs running the compute load, N( _Ac ):

N( _Ac )=16-N( _Ab );

(7.2) Determine whether the number of physical CPUs set to 1 bit in the I/O bitmap A _b is less than N( _Ab ), then proceed to sub-step (7.3), otherwise the rotor step (7.4);

(7.3) Transfer the number of physical CPUs from the physical CPU running the computational load to run the I/O type load: Let Y=N( _Ab )—I/O Bitmap A _b Set the number of physical CPUs in _Ab Set the first Y bits of the I/O bitmap A _b to 1 and clear the corresponding bit in the calculated bitmap A _c to the rotor step (7.6);

(7.4) Determine whether the number of physical CPUs set to 1 in the I/O bitmap A _b is equal to N( _Ab ), then the rotor step (7.7); otherwise, perform sub-step (7.5);

(7.5) Transfer the number of physical CPUs from the physical CPU running the I/O type load to run the computational load: Let Y=I/O bitmap A _b set the number of physical CPUs - N( _Ab ) , clear the first Y bits in the I/O bitmap A _b to 0, and calculate the corresponding position in the bitmap A _c , the rotor step (7.6);

(7.6) reset the virtual CPU affinity: reestablish all virtual CPU and I Run I / O type load / O bit map A _b is set in a physical affinity for the CPU;

Re-establish the operating computational load is calculated for all virtual CPU and bit map A _c is set in a physical affinity for the CPU;

(7.7) Clear the I/O load variable L _b and the calculated load variable L _{c to} 0, and update the I/O task urgency variable I _{ij of} all virtual CPUs: assign the value of 1/H _ij to I _ij , the rotor step ( 3.4);

(8) As shown in FIG. 6, the step of changing the virtual machine type includes the following substeps:

(8.1) According to the modified virtual machine type command issued by the user, the virtual machine ID, the virtual machine type Q, and the virtual CPU field X _i are read, and the virtual machine type is read in the virtual machine scheduling structure according to the virtual machine ID. S _i ;

(8.2) judging whether Q=0, yes, proceeding to sub-step (8.3), otherwise performing sub-step (8.6);

(8.3) Determine whether S _i =0, if yes, keep the virtual machine type S _i unchanged, the rotor step (3.5); otherwise, judge whether S _i =1, then proceed to sub-step (8.4), otherwise perform sub-step (8.5 );

(8.4) Determining that S _i =1 means that the virtual machine running the I/O type load is converted into a virtual machine of the computational load, and the physical CPUs of each virtual CPU and the I/O bitmap A _b are removed. affinity, to re-establish the virtual machine and the virtual CPU is calculated for all bit map a physical CPU affinity of a center _c, and the modified S _i is 0, the rotor step (3.5);

(8.5) Determining S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the computational load, and the virtual CPUs of each virtual CPU and the I/O bitmap A _b are deactivated. and resistance, the re-establishment of the virtual CPU which calculates the respective bit map a _c is set in a physical CPU affinity, and S _i 0 is modified, the rotor step (3.5);

(8.6) judging whether Q=1, if yes, proceeding to sub-step (8.7), otherwise performing sub-step (8.11);

(8.7) judging whether S _i =0, if yes, proceeding to step (8.8), otherwise performing sub-step (8.9);

(8.8) indicates that the virtual machine running the computational load is converted into a virtual machine running an I/O type load, and the affinity of each virtual CPU to the physical CPU that is set to one bit in the calculation bitmap A _c is released, and the All virtual CPUs of the virtual machine have affinity with the physical CPU set to 1 bit in the I/O bitmap A _b , and modify S _i to 1, the rotor step (3.5);

(8.9) judging whether S _i =1, yes, keeping the virtual machine type S _i unchanged, the rotor step (3.5); otherwise performing the sub-step (8.10);

(8.10) Judging that S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the I/O type load, and the virtual CPU of each of the virtual CPUs is calculated and the physical CPU of the bitmap A _c is set. And re-establish the affinity of each virtual CPU with the physical CPU set to 1 bit in the I/O bitmap A _i , and modify S _i to 1, the rotor step (3.5);

(8.11) Determine Q=2, judge whether S _i =0, if yes, perform sub-step (8.12), otherwise perform sub-step (8.13);

(8.12) Determining S _i =0, indicating that the virtual machine running the computational load is converted into a virtual machine running the hybrid load, and the first X _i virtual CPUs in the virtual machine are removed and the calculation bitmap A _c is set to 1 bit. The affinity of the physical CPU, re-establishing the affinity of the first X _i virtual CPUs in the virtual machine and the physical CPU set in the I/O bitmap A _b , and modifying the S _i to 2, the rotor step ( 3.5);

(8.13) Judging whether S _i =1, then proceeding to sub-step (8.14), otherwise performing sub-step (8.15);

(8.14) determines S _i = 1, represents the operation I / O type load into the virtual machine a virtual machine running the hybrid load is released after the virtual machine M _i -X _i virtual CPU and I / O bitmap The affinity of the physical CPU of the 1-bit is set in A _b , and the affinity of the virtual CPU in the virtual machine after the M _i -X _i virtual CPU is calculated and the physical CPU set in the bitmap A _c is set, and the S is _{i is} modified to 2, rotor step (3.5);

(8.15) Determine S _i = 2, keep the virtual machine type S _i unchanged, and rotate the step (3.5).

In this embodiment, in the steps of scheduling initialization, reallocating the physical CPU, and changing the virtual machine type, establishing each virtual CPU and the I/O bitmap A _b or calculating the physical CPU of the bitmap A _c Affinity, through the system-defined setting affinity function int vcpu_set_affinity (struct vcpu * v, const cpumask_t *affinity), where the parameter struct vcpu * v is the virtual CPU pointer that needs to set the affinity, cpumask_t *affinity is set Affinity target physical CPU bitmap pointer;

When the virtual CPU is deactivated with the I/O bitmap A _b or the physical CPU of the bitmap A _c , the system-defined setting affinity function int vcpu_set_affinity(struct vcpu*v, const cpumask_t *affinity), re-establish the virtual CPU and I / O bitmap A _b or calculate the affinity of the physical CPU set in the bitmap A _c , where the parameter struct vcpu * v is the virtual need to reset the affinity CPU pointer, cpumask_t*affinity is the target physical CPU bitmap pointer for setting affinity.

Claims (3)

1. A virtual CPU scheduling method includes a scheduling initialization step, a virtual CPU running queue step, a virtual CPU running step, a load balancing step, an update credit value and system load step, a process of reallocating a credit value, reallocating a physical CPU step, and changing a virtual Machine type step, which is characterized by:

   (1) Scheduling initialization step: Create a physical CPU scheduling structure for each physical CPU: set a main timer and an adjustment timer for the main physical CPU, and set a secondary timer for other physical CPUs; create a virtual machine scheduling structure for each virtual machine : Initialize the virtual machine credit value weight, set the virtual machine type; initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU: initialize the load variable of each virtual CPU, the I/O task urgency, and allocate for each virtual CPU. Credit value; initialize I/O bitmap A _b and calculate bitmap A _c , the physical CPU in A _b is dedicated to scheduling virtual CPU running I / O type load, the physical CPU in A _c is dedicated to scheduling running computing load Virtual CPUs, which are complementary physical CPU bitmaps; initialize the global I/O load variable L _b and calculate the load variables L _c , L _b and L _c to record the I/O load and the calculated load in the system, respectively. The credit value resource consumed by the virtual CPU; setting the priority of each virtual CPU;

   (2) Insert virtual CPU running queue step: If the current virtual CPU runs the computing load, it is directly inserted after all virtual CPUs of the same priority in the virtual CPU running queue; if the current virtual CPU runs the I/O type load, then press it The priority is sequentially inserted into the virtual CPU running queue. For virtual CPUs with the same priority, the I/O urgency is sorted from large to small.

   (3) Step of running the virtual CPU: If the credit value of the virtual CPU of the current physical CPU is less than or equal to 0, trigger load balancing, and go to step (4); otherwise, run the virtual CPU of the team first, if the current physical CPU is the primary physical CPU. During the running process, the secondary timer reaches the time, the step (5), the main timer to the time, the step (6), the timer is adjusted to the time, the step (7) is performed, and the user issues a command to modify the virtual machine type. Step (8) is performed; the current physical CPU is a non-primary physical CPU. During the running process, the secondary timer expires, and the step (5) is performed. The user issues a command to modify the virtual machine type, and the step (8) is executed; the virtual CPU reaches the running. Turn after time Step (2), reinserting the queue;

   (4) Load balancing step: determine whether the corresponding bit of the current physical CPU in the I/O bitmap is 1, and the non-idle physical CPU with the I/O bitmap set to 1 bit as the target CPU group, otherwise the bitmap will be calculated. Set a non-idle physical CPU as the target CPU group; then select the optimal physical CPU from the target physical CPU group as the target physical CPU of the migration, and abandon the load balancing if the optimal physical CPU is not found;

   After the target physical CPU is found, it is determined whether the current physical CPU has a corresponding bit in the I/O bitmap, and the virtual CPU that is the most urgent I/O task in the virtual CPU running queue of the target physical CPU is used as the target virtual CPU. Otherwise, the virtual CPU in the virtual CPU running queue of the target physical CPU that has the historical load closest to the difference between the target physical CPU and the current physical CPU historical load is taken as the target virtual CPU, and the target virtual CPU is deleted from the original queue and inserted into the current physical CPU. Virtual CPU queue leader; turn back to step (3);

   (5) Update credit value and system load step: update the credit value and load variable of the current virtual CPU, and update the historical load variable of the relevant physical CPU, and determine the corresponding physical value of the physical CPU running the current virtual CPU in the I/O bitmap. Whether it is 1, is to update its I / O task pressing variables and global I / O load variables, go back to step (3); otherwise update the global computing load variables; go back to step (3);

   (6) Re-allocating the credit value step: determining the priority according to the credit value of the current virtual CPU, recalculating the credit value quota allocated by each virtual CPU, and updating the credit value of all the virtual CPUs to determine whether the credit value of the updated virtual CPU is Change from non-positive to positive, then change its priority, and return to step (3); otherwise, the priority of the virtual CPU does not change, and directly returns to step (3);

   (7) Reassign the physical CPU step: Calculate the number of physical CPUs running the I/O load and the number of physical CPUs running the calculated load according to the global I/O load variable and the calculated load variable; and modify according to the calculated physical CPU number Corresponding I/O bitmap and calculation bitmap, re-establish the affinity between virtual CPU and physical CPU; clear I/O load variable and calculation load variable to 0, and update all virtual CPU I/O tasks urgently Degree variable, return to step (3);

   (8) to change the type of virtual machine steps of: reading a user command, according to which the virtual your crash ID, a virtual machine type and number of virtual Q CPU X _i, read the original virtual machine type S _i, according to the original virtual machine and a virtual machine type Q Type S _i , re-establish the affinity of each virtual CPU of the virtual machine with the I/O bitmap or calculate the bitmap, and modify the virtual machine type variable, and return to step (3).
2. The virtual CPU scheduling method according to claim 1, wherein:

   (1) The scheduling initialization step includes the following sub-steps:

   (1.1) Create a physical CPU scheduling structure for each physical CPU:

   Determine whether to create a physical CPU scheduling structure for the first physical CPU, and then perform sub-step (1.2); otherwise, the rotor step (1.3);

   (1.2) Mark the first physical CPU as the main physical CPU, and set the main timer and the adjustment timer for it. The main timer period T _main = 30 ~ 100ms, which is an integer multiple of 10ms, and is an integer of T _times . times, adjusting the timer period T _modulated = 1200 ~ 3000ms, T is an integral multiple of _{the main} rotor step (1.3);

   (1.3) a secondary set timer period of the timer T _times as the secondary (1 ~ 3) × 10ms; the physical CPU load history variable is initialized to 0 H _p; H _p variable load history recorded on the physical CPU The sum of all the credits consumed by the running virtual CPU, the physical CPU number p = 1 ~ G, G is the number of physical CPUs in the system, G = 1 ~ 64;

   (1.4) Create a virtual machine scheduling structure for each virtual machine:

   In the virtual machine scheduling structure of each virtual machine, the corresponding virtual machine credit value weight W _{i is} initialized to 256, the corresponding virtual machine type S _{i is} set to 0, and its virtual CPU linked list is initialized to be empty;

   The virtual machine number i=1～V, V is the number of virtual machines created by the user, V=1～100; S _i is 0 indicates a computing load, and S _i is 1 indicates an I/O type load, and S _i is 2 indicates a hybrid load, and the hybrid load is a load equivalent to a calculation type and an I/O type load;

   (1.5) Initialize the virtual CPU scheduling structure for each virtual machine's virtual CPU:

   Add the virtual CPU to the virtual CPU linked list of the virtual machine to which it belongs;

   Initialize the load variable H _{ij of} each virtual CPU to 0,

   Initialize the I/O task urgency variable I _{ij of} each virtual CPU to 0, I _ij =0~1. The larger the I _ij value, the more urgent the I/O task is, I _ij =0, indicating no I/O load. ;

   And assigning a credit value to each virtual CPU, wherein the credit value of the jth virtual CPU of the i-th virtual machine is C _ij = G × 10 × T _main × W _i / (W _T × M _i );

   Among them, the total weight of the system

   

   The number of virtual CPUs in the virtual machine is j=1～M _i , and M _i is the number of virtual CPUs of the i-th virtual machine, which is set by the user and is an integer smaller than G;

   (1.6) Initialize the I/O bitmap A _b and the calculated bitmap A _c :

   Establish I / O bit map bit map calculation and A _b A _c, A _b is the bits are initialized to all-0, A _c is initialized to all you, a system default initialization operation computational load;

   (1.7) Upon initialization, all default virtual CPU computational load operation, establishing each of the virtual CPU calculates the bit map A _c is set a physical CPU affinity;

   (1.8) setting the global I / O load variable L _b and calculating the load variable L _c , and initializing both L _b and L _c to 0;

   (1.9) Initializing the priority P _{ij of} each virtual CPU to -1, indicating that it can be scheduled;

   (2) Insert the virtual CPU run queue step, including the following substeps:

   (2.1) determining whether the current virtual CPU has an affinity with the calculation bitmap A _c , and indicating that the virtual CPU is running a computational load, performing sub-step (2.2), otherwise indicating that the virtual CPU is running an I/O type load, the rotor Step (2.3),

   (2.2) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPUs in the queue are compared, otherwise the sequence is compared backwards until the virtual CPU runs the queue tail; after the insertion is completed, go to step (3);

   (2.3) Starting from the virtual CPU running queue head of the current physical CPU, comparing whether the current virtual CPU priority P _ij is greater than the virtual CPU priority P _ij in the virtual CPU running queue, and inserting the current virtual CPU into the virtual CPU. Before the virtual CPU is compared in the queue, go to step (3); otherwise, perform sub-step (2.4);

   (2.4) Compare whether the priority P _{ij of the} current virtual CPU is equal to the virtual CPU priority P _ij in the virtual CPU running queue, and then perform sub-step (2.5), otherwise the sequence is compared backwards until the virtual CPU runs the queue tail. After the insertion is completed, go to step (3);

   (2.5) comparing the current value of the virtual CPU if I _ij I _ij is greater than the value of the virtual CPU operation queue virtual CPU, and a transfer step (3) before insertion then the current virtual CPU virtual CPU compares the run queue virtual CPU; Otherwise, the order is backward, comparing whether the priority P _{ij of the} current virtual CPU is greater than the virtual CPU priority P _ij for comparison, and then inserting the current virtual CPU into the virtual CPU running queue to compare the virtual CPUs and then transferring to step (3); Otherwise sub-step (2.4);

   (3) The virtual CPU running steps include the following sub-steps:

   (3.1) In the virtual CPU running queue of the current physical CPU, determine whether the team head virtual CPU credit value C _ij is greater than 0, then proceed to sub-step (3.2); otherwise, trigger the load balancing policy, and then go to step (4);

   (3.2) Determine whether the target physical CPU group set by the first virtual CPU affinity of the virtual CPU running queue is the physical CPU represented by the 1 bit in the A _b bitmap, indicating that it is running the I/O type load and setting it The running time slice length T _s =1ms ~ 5ms; otherwise it indicates that it runs the computational load, set its running time slice length T _s = 30ms ~ 100ms;

   (3.3) deleting the team head virtual CPU from the virtual CPU running queue;

   (3.4) determining whether the current physical CPU is the primary physical CPU, if yes, performing sub-step (3.5); otherwise, performing sub-step (3.8);

   (3.5) running the team head virtual CPU on the primary physical CPU, the secondary timer of the primary physical CPU, the primary timer, and the adjustment timer are simultaneously clocked; and simultaneously monitoring whether the user issues a modified virtual If the machine type command is executed, step (8) is performed; otherwise, sub-step (3.6) is performed;

   (3.6) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.7); otherwise, removing the first virtual CPU from the primary physical CPU, and then moving to step (2);

   (3.7) When the secondary timer has experienced the time T _times , the secondary timer is reset, go to step (5); when the master timer has experienced the time T _master , the master timer is reset, go to step (6); The timer has experienced the time T _adjustment , then adjust the timer reset, go to step (7); if no timer is triggered within T _s time, the virtual CPU continues to run for T _s , the team first virtual CPU from the main physical CPU Remove, go to step (2);

   (3.8) Running the team head virtual CPU on the current physical CPU, and timing the secondary timer of the current physical CPU; and monitoring whether the user modifies the virtual machine type, if yes, performing step (8); otherwise, performing sub-step (3.9) ;

   (3.9) determining whether the running time of the first virtual CPU of the team is less than T _s , then proceeding to sub-step (3.10); otherwise, removing the first virtual CPU from the current physical CPU, and proceeding to step (2);

   (3.10) when subjected to a secondary timer _times the time T, then the secondary timer is reset, go to step (5), if the timer does not trigger any time T _S, the virtual CPU operation time duration T _S, the team The first virtual CPU is removed from the current physical CPU, and the process proceeds to step (2);

   (4) The load balancing step includes the following substeps:

   (4.1) Select the physical CPU group for load balancing: determine whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, indicating that the virtual CPU running queue of the physical CPU is running I/O type. Load, the non-idle physical CPU set to 1 bit in the I/O bitmap A _b is used as the target physical CPU group for load balancing, and sub-step (4.2); otherwise, the physical CPU queue is running the computational load. Calculating a non-idle physical CPU set to 1 bit in the bitmap A _c as a target physical CPU group for load balancing, and performing sub-step (4.2);

   (4.2) Selecting the optimal physical CPU from the target physical CPU group: traversing the target physical CPU group in the corresponding A _b or A _c , and determining whether to find the non-physical CPU on the same physical core The idle physical CPU is the target physical CPU of the migration, the rotor step (4.8), otherwise the rotor step (4.3);

   (4.3) Re-traversing the target physical CPU group in the corresponding A _b or A _c to determine whether a non-idle physical CPU in the same physical package as the current physical CPU is found, and then sub-step (4.4), otherwise the rotor Step (4.5);

   If (4.4) determines the non-physical CPU idle is greater than the current H _p H _p physical CPU, the physical CPU is the target physical CPU, the rotor step (4.8) of the migration, otherwise step rotor (4.5);

   (4.5) in the corresponding target physical CPU group, sequentially determine backwards whether to find a non-idle physical CPU in the same physical package as the current physical CPU, then the rotor step (4.4), otherwise the rotor step (4.6);

   (4.6) Re-traversing the target physical CPU group to determine whether to find a non-idle physical CPU on a different physical package than the current physical CPU. If yes, perform sub-step (4.7). Otherwise, mark the current physical CPU as idle and run idle virtual. CPU, rotor step (3.4);

   (4.7) determining whether the H _{p of} the non-idle physical CPU is greater than twice the H _p of the current physical CPU, and the non-idle physical CPU is the target physical CPU of the migration, the rotor step (4.8); otherwise, the current physical CPU Marked as idle, running idle virtual CPU, rotor step (3.4);

   (4.8) Select the appropriate virtual CPU to transfer on the target physical CPU:

   Determine whether the corresponding bit of the current physical CPU in the I/O bitmap A _b is 1, and traverse the virtual CPU running queue of the target physical CPU, and use the virtual CPU with the largest I _ij value as the target virtual CPU of the migration, and the rotor step ( 4.9); H _p half of the difference between the target virtual CPU or physical CPU traverse the virtual CPU run queue, find H _ij closest to the target physical CPU H _p and trigger load balancing physical CPU as the migration target virtual CPU, rotor step (4.9);

   (4.9) removing the migrated target virtual CPU from the virtual CPU running queue of the target physical CPU and inserting it into the head of the virtual CPU running queue of the current physical CPU, and the rotor step (3.2);

   (5) The steps of updating the credit value and system load, including the following sub-steps:

   (5.1) updating the credit value C _{ij of the} current virtual CPU: assigning the value of C _ij -10×T ₁ to C _ij ;

   Update the load variable H _{ij of the} current virtual CPU: assign the value of H _ij +10×T ₁ to H _ij ;

   Update the historical load variable H _p of the physical CPU running the current virtual CPU: assign the value of H _p +10×T ₁ to H _p ;

   The credit value of a virtual CPU running for 1ms is 10, and T ₁ is the time elapsed since the last virtual CPU has been running since the last time, in units of ms;

   (5.2) Determine whether the physical CPU running the current virtual CPU has a corresponding bit value in the I/O bitmap A _b , which means that the I/O task is run, and its I _{ij is} updated, and I _ij -1/(10 The value of ×T ₁ ) is assigned to I _ij , and sub-step (5.3) is performed; otherwise, sub-step (5.3) is directly performed;

   (5.3) According to the current type of virtual CPU running load, the credit value consumed by the virtual CPU is added to the system load: it is determined whether the physical CPU running the current virtual CPU has a corresponding bit value in the bitmap A _b , and the global value is updated. The I/O load variable L _b , the value of L _b +10×T ₁ is assigned to L _b , the rotor step (3.4); otherwise the global calculated load variable L _{c is} updated, and the value of L _c +10×T ₁ is assigned L _c , rotor step (3.4);

   (6) The step of reallocating the credit value includes the following substeps:

   (6.1) determining whether the current virtual CPU credit value is greater than 0, then proceeding to sub-step (6.2); otherwise, setting the current virtual CPU priority P _ij to -2, indicating that the scheduling cannot be performed again, performing sub-step (6.2);

   (6.2) Newly assign a credit value NC _{ij to} the virtual CPU of each virtual machine:

   NC _ij = G × T _main × 10 × W _i / (W _T × M _i );

   Where NC _ij is the newly obtained credit value of the jth virtual CPU of the i-th virtual machine;

   (6.3) Update the credit value C _{ij of} each virtual CPU:

   Assign the value of NC _ij +C _ij to C _ij ,

   Judging whether C _ij >0 and P _ij value is -2, is to change the priority P _ij about the virtual CPU to -1, the rotor step (3.4), otherwise the priority P _ij about the virtual CPU does not change, the direct rotor Step (3.4);

   (7) The process of reallocating the physical CPU includes the following substeps:

   (7.1) Calculate the number of corresponding physical CPUs;

   Number of physical CPUs running I/O type load N( _Ab ):

   N( _Ab )=(L _b /(L _b +L _c ))×G;

   The number of physical CPUs running the compute load, N( _Ac ):

   N( _Ac )=G-N( _Ab );

   (7.2) Determine whether the number of physical CPUs set to 1 bit in the I/O bitmap A _b is less than N( _Ab ), then proceed to sub-step (7.3), otherwise the rotor step (7.4);

   (7.3) Transfer the number of physical CPUs from the physical CPU running the computational load to run the I/O type load: Let Y=N( _Ab )—I/O Bitmap A _b Set the number of physical CPUs in _Ab Set the first Y bits of the I/O bitmap A _b to 1 and clear the corresponding bit in the calculated bitmap A _c to the rotor step (7.6);

   (7.4) Determine whether the number of physical CPUs set to 1 in the I/O bitmap A _b is equal to N( _Ab ), then the rotor step (7.7); otherwise, perform sub-step (7.5);

   (7.5) Transfer the number of physical CPUs from the physical CPU running the I/O type load to run the computational load: Let Y=I/O bitmap A _b set the number of physical CPUs - N( _Ab ) , clear the first Y bits in the I/O bitmap A _b to 0, and calculate the corresponding position in the bitmap A _c , the rotor step (7.6);

   (7.6) reset the virtual CPU affinity: reestablish all virtual CPU and I Run I / O type load / O bit map A _b is set in a physical affinity for the CPU;

   Re-establish the operating computational load is calculated for all virtual CPU and bit map A _c is set in a physical affinity for the CPU;

   (7.7) Clear the I/O load variable L _b and the calculated load variable L _{c to} 0, and update the I/O task urgency variable I _{ij of} all virtual CPUs: assign the value of 1/H _ij to I _ij , the rotor step ( 3.4);

   (8) Steps to change the virtual machine type, including the following sub-steps:

   (8.1) According to the modified virtual machine type command issued by the user, read the virtual machine ID, the virtual machine type Q, and the virtual CPU number X _i , and read the virtual machine type in the virtual machine scheduling structure according to the virtual machine ID. S _i ;

   The modify virtual machine type command includes a virtual machine ID field, a virtual machine type field, and a virtual CPU number field, and the virtual machine ID field gives a virtual machine ID of a type to be modified; the virtual machine type field gives a virtual machine type Q to be modified. Q is 0 for the computational load, Q is 1 for the I/O load, and Q is 2 for the hybrid load; the virtual CPU number field is for the number of virtual CPUs used to run the I/O load in the hybrid virtual machine. This field is valid only when Q=2;

   (8.2) judging whether Q=0, yes, proceeding to sub-step (8.3), otherwise performing sub-step (8.6);

   (8.3) Determine whether S _i =0, if yes, keep the virtual machine type S _i unchanged, the rotor step (3.5); otherwise, judge whether S _i =1, then proceed to sub-step (8.4), otherwise perform sub-step (8.5 );

   (8.4) Determining that S _i =1 means that the virtual machine running the I/O type load is converted into a virtual machine of the computational load, and the physical CPUs of each virtual CPU and the I/O bitmap A _b are removed. affinity, to re-establish the virtual machine and the virtual CPU is calculated for all bit map a physical CPU affinity of a center _c, and the modified S _i is 0, the rotor step (3.5);

   (8.5) Determining S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the computational load, and the virtual CPUs of each virtual CPU and the I/O bitmap A _b are deactivated. and resistance, the re-establishment of the virtual CPU which calculates the respective bit map a _c is set in a physical CPU affinity, and S _i 0 is modified, the rotor step (3.5);

   (8.6) judging whether Q=1, if yes, proceeding to sub-step (8.7), otherwise performing sub-step (8.11);

   (8.7) judging whether S _i =0, if yes, proceeding to step (8.8), otherwise performing sub-step (8.9);

   (8.8) indicates that the virtual machine running the computational load is converted into a virtual machine running an I/O type load, and the affinity of each virtual CPU to the physical CPU that is set to one bit in the calculation bitmap A _c is released, and the All virtual CPUs of the virtual machine have affinity with the physical CPU set to 1 bit in the I/O bitmap A _b , and modify S _i to 1, the rotor step (3.5);

   (8.9) judging whether S _i =1, yes, keeping the virtual machine type S _i unchanged, the rotor step (3.5); otherwise performing the sub-step (8.10);

   (8.10) Judging that S _i = 2 means that the virtual machine running the hybrid load is converted into a virtual machine running the I/O type load, and the virtual CPU of each of the virtual CPUs is calculated and the physical CPU of the bitmap A _c is set. And re-establish the affinity of each virtual CPU with the physical CPU set to 1 bit in the I/O bitmap A _i , and modify S _i to 1, the rotor step (3.5);

   (8.11) Determine Q=2, judge whether S _i =0, if yes, perform sub-step (8.12), otherwise perform sub-step (8.13);

   (8.12) Determining S _i =0, indicating that the virtual machine running the computational load is converted into a virtual machine running the hybrid load, and the first X _i virtual CPUs in the virtual machine are removed and the calculation bitmap A _c is set to 1 bit. The affinity of the physical CPU, re-establishing the affinity of the first X _i virtual CPUs in the virtual machine and the physical CPU set in the I/O bitmap A _b , and modifying the S _i to 2, the rotor step ( 3.5);

   (8.13) Judging whether S _i =1, then proceeding to sub-step (8.14), otherwise performing sub-step (8.15);

   (8.14) determines S _i = 1, represents the operation I / O type load into the virtual machine a virtual machine running the hybrid load is released after the virtual machine M _i -X _i virtual CPU and I / O bitmap The affinity of the physical CPU of the 1-bit is set in A _b , and the affinity of the virtual CPU in the virtual machine after the M _i -X _i virtual CPU is calculated and the physical CPU set in the bitmap A _c is set, and the S is _{i is} modified to 2, rotor step (3.5);

   (8.15) Determine S _i = 2, keep the virtual machine type S _i unchanged, and rotate the step (3.5).
3. The virtual CPU scheduling method according to claim 2, wherein:

   In the steps of the scheduling initialization step, the reallocation of the physical CPU step, and the change of the virtual machine type, when each virtual CPU is established with the I/O bitmap A _b or the affinity of the physical CPU set in the bitmap A _c is set, Set the affinity function through the system definition;

   When the virtual CPU is deactivated with the I/O bitmap A _b or the physical CPU that is set to one bit in the bitmap A _c , the virtual CPU and I/O are re-established by the system-defined setting affinity function. The bitmap A _b or the affinity of the physical CPU set to 1 bit in the bitmap A _c is calculated.

Images