System and method for evaluating a workload and its impact on performance of a workload manager

• US 20080022285A1
• Filed: 07/20/2006
• Published: 01/24/2008
• Est. Priority Date: 07/20/2006
• Status: Active Grant

• Current Assignee: Hewlett Packard Enterprise Development LP (Hewlett-Packard Enterprise Company)
• Original Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Inventors: McCarthy, Clifford A., Rolia, Jerome, Cherkasova, Ludmila
• Granted Patent: US 8,112,756 B2
• Time in Patent Office: Days
• US Class Current: 718/104
• CPC Class Codes: G06F 2209/5022 Workload threshold
• CPC Class Codes: G06F 9/505 considering the load

Abstract

A system comprises a workload evaluator that is operable to receive a representative workload that is representative of competing demands for capacity of at least one shared computing resource. The workload evaluator evaluates the representative workload and computes a metric representing a degree of burstiness of demands present in the representative workload. The metric representing degree of burstiness of the representative workload may be used for estimating an upper bound on quality of service provided by a workload manager to the representative workload. The metric may also be used for evaluating at least one scheduler parameter setting of the workload manager to aid in determining an optimal parameter setting based at least in part on the estimated impact of the representative workload on QoS provided by the workload manager.

Claims

1. A method comprising:

receiving, by a workload evaluator, a representative workload that comprises a plurality of competing demands for capacity of at least one shared resource;

evaluating, by the workload evaluator, the representative workload to determine at least one characteristic of the representative workload;

determining, based at least in part on the at least one characteristic of the workload, an impact of the representative workload on performance of a workload manager in managing access of the representative workload to the capacity of the at least one shared resource.

2. The method of claim 1 wherein the at least one characteristic is a characteristic that impacts performance of a workload manager.

3. The method of claim 1 wherein the at least one characteristic comprises a degree of burstiness of the representative workload.

4. The method of claim 3 further comprising:

computing, by the workload manager, a metric (B^D) representing the degree of burstiness as follows;

$B^D=\frac{\underset{i=0}{\overset{i=T}{\sum <br><br>}}<br><br>D_i}{D}<br><br><br><br>\mathrm{such}<br><br><br><br>\mathrm{that}<br><br><br><br>\frac{\max <br><br>\left(\mathrm{Demand\_minResource},D_i\right)}{\max <br><br>\left(\mathrm{Demand\_minResource},D_{i-1}\right)}\le <br><br>\mathrm{demand\_variation},$

wherein the representative workload comprises resource access demands for intervals i=1, . . . , T, wherein D_i represents demand for access to the shared resource for interval i, wherein Demand_minResource is a threshold defining a boundary for demands of interest in the representative workload, and wherein demand-variation defines a limit on a ratio in successive demands present in the representative workload that are of interest.

5. The method of claim 4 further comprising:

defining Demand_minResource as Demand_minResource=minResource allocation/demand_variation, wherein minResource allocation is a scheduler control parameter that defines a minimum allocation of the at least one shared resource for a respective consumer during any scheduling interval.

6. The method of claim 4 further comprising:

defining

$\mathrm{demand\_variation}<br><br><br><br>\mathrm{as}<br><br><br><br>\mathrm{demand\_variation}=\frac{\mathrm{upperAllocUtil}}{\mathrm{lowerAllocUtil}},$

wherein lowerAllocUtil is a scheduler control parameter that triggers a decrease of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer drops below the lowerAllocUtil parameter, and wherein upperAllocUtil is a scheduler control parameter that triggers an increase of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer increases above the upperAllocUtil parameter.

7. The method of claim 1 further comprising:

evaluating at least one scheduler parameter setting of the workload manager based at least in part on the determined impact of the representative workload on performance of the workload manager.

8. The method of claim 7 wherein the at least one scheduler parameter setting comprises a lowerAllocUtil threshold parameter, wherein the lowerAllocUtil threshold parameter triggers a decrease of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer drops below the lowerAllocUtil threshold parameter.

9. The method of claim 1 wherein the determining an impact of the representative workload on performance of the workload manager comprises:

estimating, based at least in part on the at least one characteristic of the workload, an upper bound on quality of service provided by the workload manager to the representative workload.

10. Computer-executable software code stored to a computer-readable medium, the computer-executable software code comprising:

code for receiving a representative workload that comprises competing demands for capacity of at least one shared resource;

code for evaluating the representative workload to determine a metric representing a degree of burstiness of the representative workload; and

code for determining, based at least in part on the determined metric representing the degree of burstiness of the representative workload, an impact that the representative workload will have on performance of a workload manager in managing access of the representative workload to capacity of the at least one shared resource.

11. The computer-executable software code of claim 10 further comprising:

code for computing the metric (B^D) representing the degree of burstiness as follows;

$B^D=\frac{\underset{i=0}{\overset{i=T}{\sum <br><br>}}<br><br>D_i}{D}<br><br><br><br>\mathrm{such}<br><br><br><br>\mathrm{that}<br><br><br><br>\frac{\max <br><br>\left(\mathrm{Demand\_minResource},D_i\right)}{\max <br><br>\left(\mathrm{Demand\_minResource},D_{i-1}\right)}\le <br><br>\mathrm{demand\_variation},$

wherein the representative workload comprises resource access demands for intervals i=1, . . . , T, wherein D_i represents demand for access to the shared resource for interval i, wherein Demand_minResource is a threshold defining a boundary for demands of interest in the representative workload, and wherein demand-variation defines a limit on a ratio in successive demands present in the representative workload that are of interest.

12. The computer-executable software code of claim 11 further comprising:

code for defining Demand_minResource as Demand_minResource=minResource allocation/demand_variation, wherein minResource allocation is a scheduler control parameter that defines a minimum allocation of the at least one shared resource for a respective consumer during any scheduling interval.

13. The computer-executable software code of claim 12 further comprising:

code for defining

$\mathrm{demand\_variation}<br><br><br><br>\mathrm{as}<br><br><br><br>\mathrm{demand\_variation}=\frac{\mathrm{upperAllocUtil}}{\mathrm{lowerAllocUtil}},$

wherein lowerAllocUtil is a scheduler control parameter that triggers a decrease of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer drops below the lowerAllocUtil parameter, and wherein upperAllocUtil is a scheduler control parameter that triggers an increase of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer increases above the upperAllocUtil parameter.

14. The computer-executable software code of claim 10 further comprising:

code for evaluating at least one scheduler parameter setting of the workload manager based at least in part on the determined impact of the representative workload on performance of the workload manager.

15. The computer-executable software code of claim 14 wherein the at least one scheduler parameter setting comprises a lowerAllocUtil threshold parameter, wherein the lowerAllocUtil threshold parameter triggers a decrease of allocation of capacity of the at least one shared resource to a consumer when utilization of the at least one shared resource by the consumer drops below the lowerAllocUtil threshold parameter.

16. The computer-executable software code of claim 10 wherein the code for determining an impact of the representative workload on performance of the workload manager comprises:

code for estimating, based at least in part on the metric representing the degree of burstiness of the workload, an upper bound on quality of service provided by the workload manager to the representative workload.

17. The computer-executable software code of claim 16 further comprising:

code for estimating a metric Q^D that defines a percentage of overall demand of the representative workload for which sufficient access to the at least one shared resource is allocated by the workload manager to be within a desired quality of service range.

18. A system comprising:

workload evaluator operable to receive a representative workload that is representative of competing demands for capacity of at least one shared computing resource; and

said workload evaluator operable to evaluate the representative workload and compute a metric representing a degree of burstiness of demands present in the representative workload.

19. The system of claim 18 wherein said degree of burstiness represents the degree to which demands present in the representative workload vary between successive scheduling intervals.

20. The system of claim 18 wherein the at least one shared computing resource comprises at least one central processing unit (CPU).