Workload interference estimation and performance optimization

• US 10,255,113 B2
• Filed: 03/19/2014
• Issued: 04/09/2019
• Est. Priority Date: 07/26/2010
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Inventors: Govindan, Sriram, Liu, Jie, Kansal, Aman
• Primary Examiner(s): Anya, Charles E
• Application Number:
  US14/218,985
• Publication Number:
  US 20140201741A1
• Time in Patent Office: 1,847 Days
• CPC Class Codes: G06F 9/45533 Hypervisors; Virtual machine monitors
• CPC Class Codes: G06F 9/5083 Techniques for rebalancing the load in a distributed system
• CPC Class Codes: G06F 9/52 Program synchronisation; Mutual exclusion, e.g. by means of semaphores

Abstract

Architecture that facilitates the estimation of interference among workloads (e.g., virtual machines) due to sharing of a shared resource (e.g., a shared cache of a computer processor), and optimization of a desired performance objective such as power or energy use in the presence of the interference. Estimation is to the extent of interference by characterizing the nature of shared resource usage and its effect on performance. Performance optimization is accomplished using metrics based on the above estimation, or alternatively, an explicit measurement of the interference effects. Methods are employed to estimate interference on the workload'"'"'s performance with changes in availability of the shared resource or with combinations of other workloads sharing the same resource and allocating workloads to one or more physical computers or resources to workloads such that a desired performance objective is optimized. The methods can include allocating workloads on demand.

Claims

1. A system, comprising:

a hardware processor; and

a hardware memory storing instructions which, when executed by the hardware processor, cause the hardware processor to;

access a data source of control workload configurations, the data source providing interference values reflecting interference with respect to a shared physical resource when individual control workloads execute concurrently and use the shared physical resource;

identify different test workloads for execution using the shared physical resource;

perform executions of the different test workloads concurrently with the individual control workloads while sharing the shared physical resource;

based at least on the executions, identify respective matching control workloads for the different test workloads, the matching control workloads exhibiting relatively similar performance degradation to respective test workloads during the executions;

using the data source of control workload configurations and the matching control workloads, estimate energy usage impact of interference when the different test workloads that use the shared physical resource are executed in specific workload configurations; and

adjust assignments of the different test workloads to one or more physical machines that use the shared physical resource based at least on the estimated energy usage impact of the interference.

2. The system of claim 1, wherein the instructions, when executed by the hardware processor, cause the hardware processor to:

compare performance degradation of individual test workload configurations against individual control workload configurations to determine the estimated energy usage impact of the interference.

3. The system of claim 1, wherein the instructions, when executed by the hardware processor, cause the hardware processor to:

adjust the assignments by causing an individual test workload executing on a first physical machine to execute on a second physical machine.

4. The system of claim 1, wherein the instructions, when executed by the hardware processor, cause the hardware processor to:

adjust the assignments by causing two test workloads executing on different physical machines to execute on the same physical machine.

5. The system of claim 1, wherein the instructions, when executed by the hardware processor, cause the hardware processor to:

adjust the assignments by causing multiple copies of an individual test workload to execute on two or more physical machines.

6. The system of claim 1, wherein the instructions, when executed by the hardware processor, cause the hardware processor to:

create the data source of control workload configurations; and

optimize the assignments based at least on the estimated energy usage impact.

7. The system of claim 1, wherein the shared physical resource is a processor cache.

8. A method executed by a processor, the method comprising:

accessing a data source that provides control interference values reflecting interference with respect to a shared resource when different control workload configurations of concurrently-executing control workloads use the shared resource;

identifying different actual workloads for execution using the shared resource;

executing the different actual workloads concurrently with different control workloads while sharing the shared resource;

based at least on the executing, identifying respective matching control workloads for the different actual workloads that show relatively similar performance degradation;

estimating actual interference values for different actual workload configurations by identifying corresponding control configurations in the data source; and

adjusting assignments of the different actual workloads to different physical machines having instances of the shared resource based at least on the estimated actual interference values.

9. The method of claim 8, the adjusting being performed according to a performance metric reflecting rates at which individual actual workloads in the different actual workload configurations complete units of work in a given time.

10. The method of claim 8, further comprising:

estimating an individual actual interference value of an individual actual workload configuration by identifying, in the data source, a particular control interference value for a particular control workload configuration that is similar to the individual actual workload configuration; and

using the particular control interference value for the particular control workload configuration to determine an estimated interference of the individual actual workload configuration.

11. The method of claim 8, wherein the different actual workloads are different virtual machines sharing the shared resource, and the shared resource is a processor cache.

12. The method of claim 8, further comprising assigning one or more of the different actual workloads to a single physical machine to improve performance when executing the different actual workloads.

13. The method of claim 8, further comprising assigning individual actual workloads to separate physical machines to improve performance when executing the different actual workloads.

14. The method of claim 8, further comprising optimizing performance when executing the different actual workloads, the optimizing being based at least on changes in demands on the shared resource by an individual actual workload.

15. The method of claim 8, wherein each of the different control workloads is configured to use a specified level of the shared resource.

16. The method of claim 8, further comprising optimizing performance when executing the different actual workloads, the optimizing being based at least on changes in a number of actual workloads accessing the shared resource.

17. A computer-implemented workload management method executed by a processor, the method comprising:

executing different test virtual machines concurrently with control virtual machines;

based at least on the executing, identifying, for the different test virtual machines, respective matching control virtual machines that exhibit similar interference with respect to a shared resource;

using the matching control virtual machines to estimate interference of different test virtual machine configurations by accessing a data source of control virtual machine configurations and associated interference values; and

based at least on estimated interference, adjusting assignments of the different test virtual machines to different physical machines according to a performance metric.

18. The method of claim 17, the performance metric reflecting estimated work done per unit of energy by the different test virtual machines for different potential assignments to the different physical machines.

19. The method of claim 17, wherein the adjusting the assignments comprises optimizing the performance metric.

20. The method of claim 17, the shared resource being a shared processor cache, the control virtual machines being configured to use the shared processor cache at specified levels of cache usage.