Prediction-based provisioning planning for cloud environments

• US 9,363,154 B2
• Filed: 09/26/2012
• Issued: 06/07/2016
• Est. Priority Date: 09/26/2012
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Akolkar, Rahul P., Iyengar, Arun, Meng, Shicong, Rouvellou, Isabelle, Silva-Lepe, Ignacio
• Primary Examiner(s): Mejia, Anthony
• Application Number:
  US13/627,672
• Publication Number:
  US 20140089509A1
• Time in Patent Office: 1,350 Days
• Field of Search: 709/223, 709/224, 709/226
• US Class Current: 1/1
• CPC Class Codes: G06F 2209/5019 Workload prediction
• CPC Class Codes: G06F 9/5005 to service a request
• CPC Class Codes: H04L 41/147 for predicting network behaviour
• CPC Class Codes: H04L 41/5096 wherein the managed service relates to distributed or central networked applications
• CPC Class Codes: H04L 43/04 Processing captured monitoring data, e.g. for logfile generation
• CPC Class Codes: H04L 47/70 Admission control; Resource allocation
• CPC Class Codes: H04L 47/822 Collecting or measuring resource availability data
• CPC Class Codes: H04L 47/823 Prediction of resource usage
• CPC Class Codes: H04L 47/83 based on usage prediction

Abstract

Various embodiments predict performance of a system including a plurality of server tiers. In one embodiment, a first set of performance information is collected for a base allocation of computing resources across multiple server tiers in the plurality of sever tiers for a set of workloads. A set of experimental allocations of the computing resources is generated on a tier-by-tier basis. Each of the set of experimental allocations varies the computing resources allocated by the base allocation for a single server tier of the multiple server tiers. A second set of performance information associated with the single server tier for each of the set of experimental allocations is collected for a plurality of workloads. At least one performance characteristic of at least one candidate allocation of computing resources across the multiple server tiers is predicted for a given workload based on the first and second sets of performance information.

Claims

1. A method, with an information processing system, for predicting performance of a system comprising a plurality of server tiers, the method comprising:

collecting, for a base allocation of computing resources across multiple server tiers in the plurality of server tiers, a first set of performance information associated with the multiple server tiers for a set of workloads, wherein the collecting comprises applying varying workloads from the set of workloads to the base allocation, wherein the first set of performance information comprises performance information for each of the varying workloads;

generating, on a tier-by-tier basis, a set of experimental allocations of the computing resources, wherein each of the set of experimental allocations varies the computing resources allocated by the base allocation for a single server tier of the multiple server tiers;

collecting, for each of the set of experimental allocations, a second set of performance information associated with the single server tier for a plurality of workloads; and

predicting at least one performance characteristic of one or more candidate allocations of computing resources across the multiple server tiers for a given workload based on the first set of performance information and the second set of performance information.

2. The method of claim 1, further comprising:

constructing a cross-tier performance model associated with the multiple server tiers based on the first set of performance information, wherein the cross-tier performance model predicts at least one performance characteristic of the base allocation for various workloads; and

constructing a single-tier performance model for each of the single server tiers associated with the set of experimental allocations, wherein the single-tier performance model captures performance changes between the base allocation and the single server tier for each of the set of experimental allocations associated with the single server tier.

3. The method of claim 2, wherein the performance changes are captured using non-parametric regression.

4. The method of claim 2, wherein predicting the at least one performance characteristic of the one or more candidate allocations comprises:

determining at least one performance characteristic of the base allocation for the given workload based on the cross-tier performance model;

determining, for at least one server tier of the multiple server tiers associated with the candidate allocation, a delta between the performance characteristic of the base allocation and at least one performance characteristic of the server tier for the given workload; and

combining the performance characteristic of the base allocation with the delta to obtain the predicted performance characteristic of the candidate allocation.

5. The method of claim 4, further comprising:

predicting a workload at each server tier of the multiple server tiers associated with the candidate allocation;

identifying an actual workload from the given workload at each tier; and

replacing the given workload with the actual workload that has been identified.

6. The method of claim 5, wherein predicting the workload at each server tier is based on at least one of actual workloads received by each server tier and workloads observed at each server tier.

7. The method of claim 1, further comprising:

measuring performance data associated with the multiple server tiers for plurality of different workload types;

determining a set of correlations among the performance data across the different workload types; and

reducing, based on the set of correlations, a number of workloads required to collect at least one of the first performance information and the second performance information.

8. The method of claim 7, wherein the set of correlations are captured using non-parametric regression.

9. The method of claim 1, further comprising:

identifying, based on the predicting, a candidate allocation from the one or more candidate allocations that comprises a predicted performance characteristic that satisfies a performance goal specified by a user, and that is associated with a lowest monetary cost; and

notifying the user of the identified candidate allocation.

10. The method of claim 1, wherein the multiple server tiers comprise a web server tier, an application server tier, and a database server tier.

11. The method of claim 1, wherein the predicted performance characteristic is at least one of an average response time associated with a workload and a request throughput rate.

12. The method of claim 1, wherein computing resources are varied within each of the set of experimental allocations by varying at least one of a number of virtual machines and a type of virtual machines allocated for the single server tier.

13. The method of claim 1, further comprising:

determining a cost relationship between different types of requests within a least one of the set of workloads and the plurality of workloads based on mutual information between a set of performance associated with one or more server tiers of the multiple server tiers and one or more workloads; and

constructing groups of request types from requests within the one or more workloads based on the cost relationship between different types of requests, wherein a cost of requests in a same group is defined in a same linear space.

14. The method of claim 1, further comprising:

reducing a complexity of a performance prediction model used to predict the at least one performance characteristic based on combining multiple request types within one or more workloads into a given number of request groups, wherein the given number of request groups reduce at least one dimension of a workload vector used as an input to the performance prediction model, and

wherein the given number of request groups is determined by performing gradient ascent on mutual information between a set of performance associated with one or more server tiers of the multiple server tiers and one or more workloads.

15. A method, with an information processing system, for predicting performance of a system comprising a plurality of server tiers, the method comprising:

collecting, for a base allocation of computing resources across multiple server tiers in the plurality of server tiers, a first set of performance information associated with the multiple server tiers for a set of workloads;

generating, on a tier-by-tier basis, a set of experimental allocations of the computing resources, wherein each of the set of experimental allocations varies the computing resources allocated by the base allocation for a single server tier of the multiple server tiers;

collecting, for each of the set of experimental allocations, a second set of performance information associated with the single server tier for a plurality of workloads;

constructing a cross-tier performance model associated with the multiple server tiers based on the first set of performance information, wherein the cross-tier performance model predicts at least one performance characteristic of the base allocation for various workloads;

constructing a single-tier performance model for each of the single server tiers associated with the set of experimental allocations, wherein the single-tier performance model captures performance changes between the base allocation and the single server tier for each of the set of experimental allocations associated with the single server tier; and

predicting at least one performance characteristic of one or more candidate allocations of computing resources across the multiple server tiers for a given workload based on the first set of performance information and the second set of performance information.

16. The method of claim 15, wherein the performance changes are captured using non-parametric regression.

17. The method of claim 15, wherein predicting the at least one performance characteristic of the one or more candidate allocations comprises:

determining at least one performance characteristic of the base allocation for the given workload based on the cross-tier performance model;

determining, for at least one server tier of the multiple server tiers associated with the candidate allocation, a delta between the performance characteristic of the base allocation and at least one performance characteristic of the server tier for the given workload; and

combining the performance characteristic of the base allocation with the delta to obtain the predicted performance characteristic of the candidate allocation.

18. The method of claim 15, further comprising:

predicting a workload at each server tier of the multiple server tiers associated with the candidate allocation;

identifying an actual workload from the given workload at each tier; and

replacing the given workload with the actual workload that has been identified.

19. The method of claim 15, further comprising:

measuring performance data associated with the multiple server tiers for plurality of different workload types;

determining a set of correlations among the performance data across the different workload types; and

reducing, based on the set of correlations, a number of workloads required to collect at least one of the first performance information and the second performance information.

20. The method of claim 15, further comprising:

identifying, based on the predicting, a candidate allocation from the one or more candidate allocations that comprises a predicted performance characteristic that satisfies a performance goal specified by a user, and that is associated with a lowest monetary cost; and

notifying the user of the identified candidate allocation.