KR101465657B1 - Method for Coordinated Scheduling For virtual machine - Google Patents

Abstract

The present invention provides a method for coordinated scheduling of a hypervisor for a virtual machine, capable of managing a plurality of virtual machines including an operating system to drive a plurality of virtual central processing units (vCPU). The method includes the steps of: (a) allocating virtual processors to a physical processor; (b) determining a coordination state between multithreaded workloads based on inter-processor interrupt between virtual processors and selecting the thread workloads, which are determined as being coordinated thread workloads, as coordinated scheduling targets; and (c) primarily performing scheduling (Urgent vCPU First Scheduling) with respect to an urgent virtual processor among a plurality of virtual processes based on inter-processor interrupt between virtual processors. The urgent virtual processor in the step (c) is based on the threads selected as the coordinated scheduling targets in the step (b). The inter-processor interrupt includes Translation Lookaside Buffer (TLB) inter-invalidation processor interrupt. If inter-TLB processor interrupt occurs, the virtual processor received the inter-TLB processor interrupt is scheduled, thereby reducing time required for synchronization.

Description

[0001] The present invention relates to a cooperative scheduling method for a virtual machine,

The present invention relates to a cooperative scheduling method of a hypervisor for a virtual machine managing a plurality of virtual machines including an operating system for driving a plurality of virtual processors (Virtual CPU, vCPU).

In general, the operating system schedules processes exclusively assuming that they own CPU resources. However, in a virtualized environment, because there are two scheduling layers of the operating system in the virtual machine and the hypervisor of the host machine due to the way in which the virtual processors of the plurality of virtual machines share the same physical processor resources, A situation arises. Due to the semantic gap in the virtualized environment, performance may be degraded when multithreaded application programs cooperating with each other are executed in a virtual machine.

A representative problem caused by a semantic gap in a virtualized environment is that virtual processors belonging to one virtual machine compete with each other while being scheduled on one physical processor. In particular, if one of the virtual processors is performing a spin lock, the performance is degraded by preempting another virtual processor in the same virtual machine scheduled to the same physical processor without performing a meaningful operation.

To solve this problem, various collaborative scheduling schemes for scheduling virtual processors belonging to one virtual machine together with different physical processors have been proposed. Recently, a relaxed cooperative scheduling scheme and a balanced scheduling scheme have been proposed to solve the problem that the strict policy of cooperative scheduling may lower the CPU utilization of the system. However, there is a disadvantage that the techniques introduced above do not reflect the need for collaborative scheduling of workloads that are actually being performed.

Korean Registered Patent No. 10-1145144 (registered May 05, 2012); Virtual machine scheduling method and system

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and it is an object of the present invention to provide a scheduling method which improves overall performance by shortening the synchronization delay time of parallel workloads and reducing unnecessary processor time consumption.

A collaborative scheduling method of a hypervisor for a virtual machine managing a plurality of virtual machines including an operating system for driving a plurality of virtual processors (virtual CPUs, vCPUs) of the present invention for solving the above- (a) allocating the plurality of virtual processors to a physical processor (pCPU); (b) determining whether cooperation among threads of multithreaded workloads is based on interprocessor interrupts among the plurality of virtual processors, and determining whether the cooperatively determined threads are cooperative scheduling targets (C) an Urgent vCPU First Scheduling step of scheduling an emergency virtual processor among the plurality of allocated virtual processors based on an inter-processor interruption between the virtual processors; and (c) The emergency virtual processor of step (b) may be based on threads selected for cooperative scheduling.

The step (c) further includes: (c-1) inserting an Urgent Queue into a run queue of the hypervisor scheduler according to a FIFO (First Input First Out) order for each of the physical processors Step < / RTI >

The virtual machine cooperative scheduling method of the hypervisor further comprises: after step (c), (d) checking whether there is a virtual processor that can be scheduled among a plurality of virtual processors waiting in the emergency queue, ) If there is a virtual processor that can be scheduled, the schedulable virtual processor may further include preempting another virtual processor that is being executed, and executing the virtual processor that can be scheduled.

Further, the inter-processor interrupt may request the emergency virtual processor to maintain the execution until a specific event is completed.

In addition, the inter-processor interrupt may allow the emergency virtual processor to request a time to remain in the emergency state for a specified time.

In addition, the interprocessor interrupt includes interleaving between TLB (Translation Lookaside Buffer) invalidation processors, and when the interleave between the TLB processors occurs, the virtual processor that has received interleaves between the TLB processors can be scheduled have.

Also, the inter-processor interrupt includes a reschedule interprocessor interrupt, and when the reschedule interprocessor interrupt occurs, transmits the reschedule interprocessor interrupt (reschedule interprocessor interrupt) The scheduling of one virtual processor may be delayed.

In addition, the step (a) may allocate a plurality of virtual processors belonging to one virtual machine to different physical processors.

In the step (a), only when a virtual processor is allocated to each of the physical processors, at least two of the plurality of virtual processors belonging to one virtual machine are allocated to one physical processor, (Load-conscious balance scheduling).

The present invention determines cooperations between threads of multithreaded workloads and selects the cooperatively determined threads as collaborative scheduling targets to minimize delay time due to synchronization between threads .

In addition, the present invention has an effect of improving the performance of TLB invalidation applications by preferentially scheduling an emergency virtual processor among the plurality of virtual processors in the same physical processor (Urgent vCPU First Scheduling).

In addition, the present invention integrates a priority scheduling method for an emergency virtual processor and a load / balance scheduling technique, thereby enhancing the performance of a parallel workload having many communications without degrading the performance of interactive workloads requiring high responsiveness in a virtualized environment There is an effect that can be.

1 is a graph showing the total number of IPI occurrences measured during execution of each PARSEC application.
2 is a graph showing the execution time ratio in the kernel to be amplified when each PARSEC application is executed as a streamcluster.
3 is a graph showing the time spent on the spin lock of applications with greatly amplified lock spinning time.
4 is a graph illustrating a reschedule interprocessor interrupt during execution of a streamcluster in a virtual machine having eight virtual processors.
5 is a flowchart showing a cooperative scheduling method of a hypervisor for a virtual machine.
Figure 6 shows a physical processor allocation scenario in a virtual processor of balance scheduling and load or balance scheduling.
FIG. 7 is a flow chart of a process for creating a virtual machine (8-vCPU VM) having eight virtual processors for performing parallel workloads and four virtual machines having one virtual processor A virtual machine (UP VM)) is integrated into one physical machine and executed.
FIG. 8 is a diagram showing a case where one virtual machine (8-vCPU VM) having eight virtual processors for performing a parallel workload and four virtual machines (UP VM) having one virtual processor for performing sequential workloads It is a graph showing performance degradation when sequential workloads are executed independently when executed in a physical machine.
9 is a graph illustrating weighted performance improvements when performed with freqmine
10 (a) is a graph showing the performance improvement of streamcluster when Resched-Co is applied.
FIG. 10 (b) is a graph showing the results of experiments to see how much the number of blocks affecting the performance is reduced when Resched-Co is applied.

The hypervisor collaborative scheduling method for a virtual machine according to an embodiment of the present invention is a method for observing communication between virtual processors in order to grasp a cooperation request of threads that perform communication with parallel workloads. In an embodiment of the present invention, an interprocessor interrupt (IPI) may be a communication signal generated between virtual processors in order to provide cooperative scheduling at a transparent hypervisor level without modification of the operating system.

Figure 1 shows the results of a total number of IPI occurrences measured during the execution of each PARSEC application on a virtual machine with eight virtual processors and a PARSEC 2.1 bench with multithreaded workloads FIG. 1 is a graph showing the total number of inter-processor interrupts generated while each application is independently executed using eight threads in one virtual machine, and the number of virtual processors Figure 1 shows the inter-processor interrupts in an integrated multithreaded workload. Figure 1 shows that inter-processor interrupts of various types are generated for each application, and the frequency Varies.

2 is a graph showing the execution time ratio in the kernel to be amplified when each PARSEC application is executed as a streamcluster. FIG. 2 is a graph showing the relationship between the amount of time consumed in the kernel when two virtual machines compete with each other for a physical processor and the cooperative scheduling is not applied, For example, the role of interprocessor interrupts and the relationship between kernel and user-level collaboration requirements can be seen.

FIG. 2 shows a result of experiments using a CFS (Compeletely Fair Scheduler), which is a basic scheduler of Linux, and experimenting with a streamcluster having a large amount of processor consumption while generating a large amount of communication among PARSEC applications.

Referring to FIG. 2, it can be seen that the processor time consumed in the kernel is amplified up to 30 times when executed with other competitive workloads. This amplification indicates that most of the time when executed independently also occurs in applications running at the user level. Also, applications that showed significantly increased kernel time were applications that had a high number of inter-processor interrupts. This result means that collaboration at the kernel level is required for communication-intensive applications.

Also, the amplified kernel time means that TLB (Translation Lookaside Buffer) invalidation and the lock spinning function are consumed. Most operating systems use interprocessor interrupts to signal that another processor has invalidated TLB entries. A processor that sends an inter-processor interrupt sends an inter-processor interrupt for TLB invalidation to another processor for synchronization of the TLB and performs busy-waiting until a response is received from the processor. In the virtualized environment, the receiving virtual processor of the inter-processor interrupt is delayed in scheduling by the hypervisor scheduler so that the unnecessary processor consumption of the virtual processor, which sends the inter-processor interrupt, . ≪ / RTI > Therefore, interrupts between the TLB invalidation processors can be seen as a signal to indicate that there is a virtual processor to be urgently scheduled to reduce unnecessary processor consumption.

FIG. 3 is a graph showing the time spent in the spin lock of applications with greatly amplified lock spinning time, and executed using lockstat of Linux. Referring to FIG. 3, it is possible to determine exactly what lock in the kernel actually causes unnecessary processor consumption.

Referring to FIG. 3, it can be seen that the futex-queue lock is responsible for most of the amplified rock spinning. Futex is kernel-level support that provides synchronization for user-level applications, and many applications use futex to protect the waiting queue of objects that need to be synchronized. This causes excessive competition for futex-queue locks by requesting intensive queuing when there are many synchronizations. This problem does not directly cause inter-processor interrupts unlike TLB invalidation, but excessive lock spinning can be seen with reference to FIG. 1, concentrating on applications where reschedule interprocessor interrupts are frequently generated have. This is because the reschedule interprocessor interrupt is used by the local processor to notify another processor that a new thread can be scheduled. Eventually, the reschedule interprocessor interrupt informs the hypervisor scheduler that the virtual processor that caused the inter-processor interrupt is potentially acquiring the spin lock. Based on this, the hypervisor according to the embodiment of the present invention can shorten the communication delay by preventing the virtual processor that has transmitted the reschedule interprocessor interrupt from being preempted (delaying the scheduling).

A reschedule interprocessor interrupt according to an embodiment of the present invention can be used to find a user level cooperation request. Because user-level synchronization typically uses block- or spin-then-block-based functions, user-level communication from a hypervisor's perspective is likely to be accompanied by reschedule interprocessor interrupts. Therefore, the hypervisor scheduler can utilize reschedule interprocessor interrupts to identify user-level collaboration requirements.

FIG. 4 shows a flow of time during which a reschedule interprocessor interrupt is recorded during execution of a streamcluster in a virtual machine having eight virtual processors. Since this application uses a barrier-based synchronization method, it can be seen that a reschedule interprocessor interrupt occurs intensively at the synchronization point, while the synchronization point is performed using each virtual processor. Thus, when virtual processors are scheduled simultaneously as reschedule interprocessor interrupts occur, the time consumed for synchronization is reduced.

Referring to the results of FIGS. 1 to 4, a collaborative scheduling method of a hypervisor for a virtual machine that prioritizes an emergency virtual processor according to an embodiment of the present invention (Urgent vCPU First Scheduling) A plurality of virtual machines including an operating system for operating a plurality of virtual machines, a virtual CPU, and a virtual CPU (vCPU), wherein the plurality of virtual machines (Urgent vCPU First Scheduling) of the virtual processors among the virtual processors of the virtual processors of the virtual processors.

5 is a flowchart illustrating a collaborative scheduling method of a hypervisor for a virtual machine according to an embodiment of the present invention. Referring to FIG. 5, a hypervisor cooperative scheduling method for a virtual machine according to an embodiment of the present invention includes allocating (S100) the plurality of virtual processors to a physical processor; Determining whether cooperation among threads of multithreaded workloads is based on interprocessor interrupts between the plurality of virtual processors, and selecting the cooperatively determined threads as collaborative scheduling targets (S200); Selecting an emergency virtual processor among the plurality of virtual processors in an allocated physical processor based on inter-processor interrupts between virtual processors, and (S300) performing an Urgent vCPU First Scheduling operation on the selected emergency virtual processor can do.

A collaborative scheduling method of a hypervisor that prioritizes an emergency virtual processor according to an embodiment of the present invention is a proportional-share based scheduler used in an existing operating system to maintain fairness among virtual machines. And it is possible to supplement the proportional-share based scheduler. To this end, the step (S300) of preferentially executing the emergency virtual processor among the plurality of virtual processors in the allocated physical processor may include, for each of the physical processors, an Urgent Queue according to a FIFO (First Input First Out) Into the Run Queue of the scheduler. When a virtual processor enters an emergency state in response to an inter-processor interrupt, the emergency virtual processor may be inserted into the Urgent Queue and the Run Queue of the scheduler.

5, a scheduler for a hypervisor may select a virtual processor to be executed next, a virtual processor that can be scheduled without breaking the inter-virtual machine fairness among the virtual processors waiting in the Urgent Queue first, (Step S400). If there is a virtual processor that can be scheduled, the virtual processor that can be scheduled may preempt another virtual processor that is being executed to execute the virtual processor that can be scheduled (S500).

The virtual processor according to an embodiment of the present invention can request its own emergency status in an event-based and time-based manner depending on the type of inter-processor interrupt.

The inter-processor interrupt according to an embodiment of the present invention may be an event-based method, and may request the emergency virtual processor to maintain execution until a specific event is completed. Alternatively, the inter-processor interrupt may be in a time-based manner, allowing the emergency virtual processor to request a time to remain in an emergency state for a specified time.

The type of interprocessor interrupt according to an exemplary embodiment of the present invention may be a translation lookaside buffer (TLB) invalidation processor interleave and / or a reschedule interprocessor interrupt.

The TLB translation look-aside buffer interleaving according to an embodiment of the present invention is used for TLB synchronization of each CPU. When an interleave occurs between the TLB processors, interleaving between the TLB processors is performed And may schedule the received virtual processor.

A reschedule interprocessor interrupt according to an embodiment of the present invention is for informing that a new thread has become schedulable, and when the reschedule interprocessor interrupt occurs, the reschedule interprocessor interrupt And may delay the scheduling of a virtual processor that has transmitted a reschedule interprocessor interrupt.

The virtual machine scheduling technique of the hypervisor according to an embodiment of the present invention can use its own time slice called urgent tslice so that a plurality of virtual processors can process a requested task in an emergency state. Since multiple virtual machines are concurrently involved in an urgent interprocessor interrupt, the urgent tslice must be kept short to increase overall responsiveness and at the same time long enough for the virtual processor to complete a useful task within the time slice. In the virtual machine cooperative scheduling of the hypervisor according to an embodiment of the present invention, urgent allowance is used to limit the range in which the virtual processor can borrow and use its future CPU time by exchanging short-term fairness with overall efficiency Additional control values called are available. Therefore, a virtual processor can borrow urgent tslice from its future CPU time by preempting a running virtual processor if its CPU usage time is within the range of urgent allowance than the CPU time of the currently executing virtual processor .

The hypervisor's collaborative scheduling method for the virtual machine, which prioritizes the emergency virtual processor according to the embodiment of the present invention, is characterized in that the semantic gap generated in the virtual environment in the parallel communication workload Can be mitigated. However, the issue of how to allocate a virtual processor to a physical processor in consideration of the load between each physical processor in a virtual desktop environment where various workloads of various characteristics are performed can not be solved by the hypervisor virtual machine cooperative scheduling method. Among the existing methods, a simple yet effective vCPU-to-pCPU allocation algorithm in a virtual processor is balance scheduling that allocates virtual processors belonging to one virtual machine to all other physical processors. This scheme increases the probability of being scheduled and executed together by having a plurality of virtual processors execute in different physical processors, thereby reducing the synchronization delay.

However, the balance scheduling may cause a problem of delaying the synchronization time when an imbalance in load occurs between a plurality of physical processors. Since balance scheduling allocates a plurality of virtual processors belonging to one virtual machine to different physical processors without considering the overall CPU load of the system, there is a possibility that even if there is a relatively lenient physical processor, .

Figure 6 shows a physical processor allocation scenario in a virtual processor of balance scheduling and load or balance scheduling.

6A shows a first virtual machine VM1 having first through fourth virtual processors vCPU1, vCPU2, vCPU3 and vCPU4, a second virtual machine VM2 having a fifth virtual processor vCPU5, 6 is a schematic diagram of a virtual processor allocated to a highly competitive physical processor in an environment in which a third virtual machine (VM3) having a virtual processor (vCPU6) is integrated. 6A, although the load is unbalanced between the first to fourth physical processors (pCPU1, pCPU2, pCPU3, pCPU4), the first virtual machine VM1 And the second virtual processors vCPU1 and vCPU2 are allocated to the first and second physical processors pCPU1 and pCPU2, respectively.

In order to solve such a problem, according to an embodiment of the present invention, a load-conscious (virtual) scheduling method considering a virtual desktop infrastructure environment in which various characteristics of workloads are performed in addition to a hypervisor virtual machine scheduling method balance scheduling method.

The load-awareness balance scheduling method according to an embodiment of the present invention is a method in which, in order to avoid an ineffective allocation of a virtual processor to a competing physical processor, if the balance scheduling policy selectively worsens the imbalance of the overall load, Lt; RTI ID = 0.0 > virtual processors < / RTI > That is, as shown in FIG. 6 (b), only when a virtual processor is allocated to each physical processor, at least two of the plurality of virtual processors belonging to one virtual machine are allowed to be allocated to one physical processor together do.

When the load / balance scheduling scheme is applied to the scenario 6 (a), the overall load can be balanced by allowing a plurality of virtual processors belonging to one virtual machine to be allocated to the same physical processor as shown in FIG. 6 (b) . However, since the load or balance scheduling may adversely affect the synchronization delay, the emergency virtual processor according to an embodiment of the present invention may be integrated with the method of priority scheduling (Urgent vCPU First Scheduling). In accordance with an embodiment of the present invention, collaborative scheduling of a hypervisor with integrated load or balance scheduling enables communication without degrading the performance of interactive workloads that require high responsiveness in a virtual desktop infrastructure (VDI) It has the advantage of increasing the performance of many parallel workloads.

8 to 11 are experimental results of a collaborative scheduling method of a hypervisor for a virtual machine according to an embodiment of the present invention. The experiment was performed in a KVM-based virtualization environment of the Linux kernel version 3.2.0, In order to guarantee fairness, CFS group scheduling method of Linux was used. The prototype that implements the virtual machine scheduling technique of the hypervisor according to an embodiment of the present invention is installed in the Dell Poweredge R610 equipped with two Intel Xeon X5550 2.67GHz quad core processor and 24GB of RAM. The guest operating system used Ubuntu 11.04 with version 3.2.0 of the Linux kernel. To verify the efficiency of the proposed scheme, the following scheduling policies were applied.

Resched-DP according to an embodiment of the present invention allows a virtual processor that initiated a reschedule interprocessor interrupt to enter an urgent state to delay a preemption using a time-based request.

In accordance with an embodiment of the present invention, TLB-Co uses an event-based request to invoke a virtual processor that receives interleaving between TLB (Translation Lookaside Buffer) invalidation processors, .

Resched-Co according to an exemplary embodiment of the present invention uses an urgent state to schedule a virtual processor receiving a reschedule interprocessor interrupt with a virtual processor that has transmitted an IPI using a time-based request Let them enter.

Resched-DP and TLB-Co policies according to an embodiment of the present invention are intended to alleviate kernel-level competition, and Resched-Co policies are intended to alleviate user-level competition. The preemption delay, urgent tslice, and urgent allowance values, which are parameters of cooperative scheduling of the hypervisor that prioritizes the emergency virtual processor that executes the urgent virtual processor preferentially, are 500us, 500us, and 18ms, respectively, Respectively. In order to allocate the vCPU-to-pCPU, the load-conscious balance scheduling (load-conscious balance scheduling) proposed above is performed by Urgent vCPU First Scheduling (Priority Scheduling) It is used in conjunction with the hypervisor's collaborative scheduling. In addition, CFS (Baseline), which is a basic scheduler of Linux, and Balance scheduling, which implements a probabilistic cooperative scheduling, are used for comparison with a collaborative scheduling method of a hypervisor for a virtual machine according to an embodiment of the present invention . Experiments have been run at least three times for complete overlapping of concurrently running workloads.

First, we verify the efficiency of the proposed method when parallel workloads and sequential workloads are integrated into one physical machine. We used one virtual machine (8-vCPU VM) with eight virtual processors and four virtual machines (UP VM) with one virtual processor. Each of the four virtual machines has a single-threaded version of x264 workload.

FIG. 7 is a flow chart of a process for creating a virtual machine (8-vCPU VM) having eight virtual processors for performing parallel workloads and four virtual machines having one virtual processor A virtual machine (UP VM)) is integrated into one physical machine and executed.

Referring to FIG. 7, performance is degraded when balance scheduling is applied to bodytrack, canneal, facesim, ferret, fluidanimate, and vips due to inefficient vCPU-to-pCPU allocation. Load-conscious balance scheduling detects the imbalance of the processor load and allocates a virtual processor of one virtual machine (8-vCPU VM) having eight virtual processors to a virtual machine (UP) having one virtual processor VM is not allocated to the physical processor to which the virtual processor of the VM is assigned.

In addition, when the load-conscious balance scheduling and the collaborative scheduling of the hypervisor for the urgent virtual processor are integrally applied, the virtual processors which compete with each other are adjusted, Up to 80% performance improvement over load-conscious balance scheduling. In detail, Resched-DP improves performance of applications with severe wait-queue lock competition such as dedup, facesim, ferret, fluidanimate, streamcluster, and x264. TLB-Co improves performance of TLB-Co such as dedup, ferret, And contributed to improving the performance of intensive applications. As a result, it can be confirmed that the best performance is obtained when the cooperative scheduling of the hypervisor with Urgent vCPU First Scheduling is applied together with the load-conscious balance scheduling.

FIG. 8 is a diagram showing a case where one virtual machine (8-vCPU VM) having eight virtual processors for performing a parallel workload and four virtual machines (UP VM) having one virtual processor for performing sequential workloads It is a graph showing performance degradation when sequential workloads are executed independently when executed in a physical machine.

In the experiment described above, it is a graph showing degradation of contrast performance when executed independently of a virtual machine (UP VM) having a single virtual processor performing a sequential workload. In the experiment, because all five virtual machines share the same share, the UP VM with one virtual processor should not be affected by the performance of other virtual machines, but with eight virtual processors in the case of balance scheduling We see that the virtual processor of one virtual machine (8-vCPU VM) is degraded compared to other scheduling techniques by allocating the virtual processor of the virtual machine (UP VM) with one virtual processor to the assigned physical processor have. This performance degradation implies that an allocation scheme that does not take into account the load of balance scheduling can impair fairness among virtual machines.

Next, the efficiency of the proposed method is verified when the virtual machines that execute the parallel workload are integrated into one physical machine. In this experiment, two parallel workloads were run simultaneously using one virtual machine (8-vCPU VM) with two 8 virtual processors. The workload to be performed with PARSEC applications is freqmine. In order to understand the performance improvement of the two workloads, we used the method of summing up the degree of contrast improvement when each workload was performed independently.

Figure 9 is a graph illustrating weighted performance improvement when performed with freqmine in the experimental scenario described above. D is load-aware balance scheduling, d is load-balanced scheduling, Resched-DP is a load-balanced scheduling, d is load-balanced scheduling, And integration of TLB-Co.

Referring to FIG. 9, in the case of Baseline, the performance of communication-intensive applications is degraded due to processor-intensive freqmine. This problem is due to the nature of the experiment that balances the loads of all the physical processors, so that the balance scheduling is solved to a great extent and the improved performance compared to the baseline can be confirmed. Likewise, load-conscious balance scheduling is similar to balance scheduling. When collaborative scheduling of the hypervisor with Urgent vCPU First Scheduling is applied, the performance of TLB invalidation intensive applications such as dedup, ferret, and vips is further improved. As in the previous experiment, we integrated the collaborative scheduling method of the hypervisor that performs the load-conscious balance scheduling and the urgent virtual processor first scheduling (Urgent vCPU First Scheduling) even in an environment where only the parallel workload is executed. It can be confirmed that the best performance is achieved.

Below is an experiment on whether the collaborative scheduling scheme of the hypervisor, which prioritizes the emergency virtual processor (Urgent vCPU First Scheduling), effectively improves performance by reflecting the demand for user-level collaborative scheduling.

For the experiment, a streamcluster was used which intensively uses its own built-in barrier. The streamcluster uses a spin-then-block method to spin the coherent thread to perform precise synchronization between cooperating threads and to pass the processor to other processes. Therefore, if the progress of other cooperating threads is similar at the point where synchronization is performed, and if they are scheduled together, performance can be improved by reducing the number of blocks. We used a bodytrack that interfered with running streamcluster the most.

10 (a) is a graph showing the performance improvement of streamcluster when Resched-Co is applied. Resched-Co improves streamcluster performance by 8% without affecting the performance of the bodytrack when using a spin-then-block barrier. This result is higher than when using a block-based barrier because the spin-then-block barrier is more likely to benefit when cooperating threads are executed together as described above. FIG. 10 (b) is a graph showing the results of experiments to see how much the number of blocks affecting the performance is reduced when Resched-Co is applied. As you can see in the graph, we can see that cooperative threads are scheduled together during spinning in the synchronization interval, which reduces 38% block-wait times.

Experimental results show that the proposed scheduling technique can reduce the synchronization latency of parallel workloads and reduce unnecessary processor time by mitigating existing semantic gap in a virtualized cloud environment in which various workloads are integrated. .

Claims (9)

A cooperative scheduling method of a hypervisor for a virtual machine managing a plurality of virtual machines including an operating system for driving a plurality of virtual processors (Virtual CPU, vCPU)
(a) allocating the plurality of virtual processors to a physical processor (pCPU);
(b) determining whether cooperation among threads of multithreaded workloads is based on interprocessor interrupts between the plurality of virtual processors, and determining whether the thread workloads determined to be cooperated are cooperative scheduling Selecting an object;
(c) Urgent vCPU First Scheduling of the emergency virtual processor among the plurality of allocated virtual processors based on inter-processor interrupts between the virtual processors,
The emergency virtual processor of step (c) may be based on threads selected for cooperative scheduling in step (b)
Wherein the interprocessor interrupt includes a TLB (Translation Lookaside Buffer) invalidation processor interrupts, and when an inter-TLB inter-processor interrupt occurs, a hypervisor for a virtual machine for scheduling the virtual processor that has received the inter- / RTI >
The method according to claim 1,
The step (c)
(c-1) inserting an Urgent Queue into the run queue of the hypervisor according to a FIFO (First Input First Out) sequence for each physical processor Cooperative scheduling method.
3. The method of claim 2, wherein the virtual machine cooperative scheduling method of the hypervisor
After the step (c)
(d) checking whether there is a virtual processor that can be scheduled among a plurality of virtual processors waiting in the emergency queue,
(e) if there is a virtual processor that can be scheduled, the schedulable virtual processor pre-empts another virtual processor in execution and executes the virtual processor that can be scheduled A collaborative scheduling method for a hypervisor
The method according to claim 1,
The inter-processor interrupt
And requesting the emergency virtual processor to keep running until a specific event is completed.
delete delete delete The method according to claim 1,
The step (a)
A method for collaborative scheduling of a hypervisor for a virtual machine by balancing scheduling that allocates a plurality of virtual processors belonging to one virtual machine to different physical processors.
delete

Images