Method for deploying storage system resources with learning of workloads applied thereto

• US 10,013,286 B2
• Filed: 02/24/2016
• Issued: 07/03/2018
• Est. Priority Date: 02/24/2016
• Status: Active Grant

• Current Assignee: ProphetStor Data Services, Inc.
• Original Assignee: ProphetStor Data Services, Inc.
• Inventors: Chen, Wen Shyen, Huang, Ming Jen
• Primary Examiner(s): Bullock, Jr., Lewis A
• Assistant Examiner(s): Lu, Kevin X
• Application Number:
  US15/051,819
• Publication Number:
  US 20170242729A1
• Time in Patent Office: 860 Days
• Field of Search: None
• CPC Class Codes: G06F 9/50 Allocation of resources, e.g. of the central processing unit [CPU]
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/048 Fuzzy inferencing
• CPC Class Codes: H04L 41/5025 by proactively reacting to service quality change, e.g. by reconfiguration after service quality degradation or upgrade

Abstract

A method for deploying storage system resources with learning of workloads applied to a storage system is disclosed. The method utilizes state-action fuzzy rules and action-reward fuzzy rules for the storage system to learn a best dynamic deployment of resources from the parameters of workloads. No more human interference is required. It reduces work loading of the storage system administrator and enhances performance of the storage system.

Claims

1. A method for deploying storage system resources with learning of workloads applied to a storage system, comprising the steps of:

A. setting state-action fuzzy rules for deviated percentages of characteristic parameters of the storage system from SLAs (Service Level Agreement) of workloads under each scenario of a plurality of scenarios and adjustments of resources, and action-reward fuzzy rules for adjustments of resources and reward values, wherein the scenario is a specific relation between a deviated direction of the characteristic parameters and a change of a corresponding resource;

B. generating an experience matrix where entries in each row refer to reward values under a specific state, and entries in each column refer to reward values for an adjustment of at least one resource, wherein the entries in the experience matrix are initialized to zero at an initial state and changes while the reward values are updated, and the specific state is refers to a combination of deviated percentages of at least two different kinds of characteristic parameters;

C. collecting current deviated percentages of characteristic parameters from one of the workloads, and providing predicted deviated percentages of characteristic parameters for said workload in a plurality of later time points;

D. randomly choosing one scenario of a plurality of scenarios and corresponding state-action fuzzy rules and membership functions of the chosen scenario;

E. performing membership functions of the state-action fuzzy rules of the chosen scenario based on the collected deviated percentages of characteristic parameters of said workload, and performing fuzzification, fuzzy inference, and result aggregation—

to the result of the membership functions to obtain a first action range;

F. defuzzifying the first action range to have an adjusted amount for at least one resource;

G. executing the adjusted amount in the storage system for the workload;

H. processing fuzzification, fuzzy inference, and result aggregation by inputting the provided predicted percentages of characteristic parameters of said workload to chosen scenario'"'"'s membership functions of the action-reward fuzzy rules to have a reward range;

I. defuzzifying the reward range to have a deviated reward value;

J. for the rows of the experience matrix corresponding to the states of the predicted deviated percentages of characteristic parameters, searching for the maximum value in each of the rows;

K. accumulating a deviated reward value and chosen values in a previous time point from the step I as an updated reward value and replacing the entry of the experience matrix under the state of the deviated percentages of parameters and action amount of the previous time point—

with the updated reward value, wherein the chosen values are the maximum values in each of the rows obtained from the step I of the previous cycle of step C to step K;

L. repeating step C to step K until each of the reward value stays the same or varies within a limited interval over time, wherein workloads that are not selected to be processed in step C are further processed in turns while step C to step K are repeated;

M. choosing a row in the experience matrix corresponding to observed deviated percentages of characteristic parameters; and

N. executing the specific adjustment of the resources corresponding to the maximum value among the entries in the row in the storage system.

2. The method according to claim 1, wherein if there is no previous time point for the step I, the updated reward value is set to be 0.

3. The method according to claim 1, wherein the characteristic parameter comprises latency, throughput, CPU (Central Processing Unit) load, memory usage, or TOPS (Input/Output Per Second).

4. The method according to claim 1, wherein the resource comprises read SSD (Solid State Disk) cache size, write SSD cache size, CPU share, RAM (Random Access Memory) size, block IO (Input/Output) bandwidth, number of VM (Virtual Machine), or number of provisioned storage.

5. The method according to claim 1, wherein the number of time points is 4 or 5.

6. The method according to claim 1, wherein the limited interval remains the same or decreases with time, after a learning stage has passed by.

7. The method according to claim 6, wherein the learning stage is longer than 5 hours.

8. The method according to claim 6, wherein a mean value between the upper and lower bounds of the reward values which varies within the limited interval is calculated and compared with other reward values in the same row to obtain a maximum one for the adjustment of resource if the reward value in each entry does not keep the same with time.

9. The method according to claim 1, further comprising a step F1 after the step F:

F1. keeping the amount of the resources and raising an alarm if an amount of the adjustment of one resource exceeds the remaining capacity of the resource.

10. The method according to claim 1, wherein a closer value of adjustment of the resource is chosen if one action causes an amount of adjustment of one resource falls between two consequent values of adjustments of the resource.

11. The method according to claim 1, wherein setting state-action fuzzy rules in the step A is obtained by the sub-steps of:

defining a linguistic value for each characteristic parameter;

constructing membership functions to depict degrees of all linguistic value; and

constructing the state-action fuzzy rules based on the membership functions.

12. The method according to claim 1, wherein action-reward fuzzy rules in the step A is obtained by the sub-steps of:

defining a linguistic value for each range of changed percentage of resource, each changed percentage of characteristic parameters and each amount of reward value;

constructing membership functions to depict degrees of all ranges of the deviated percentage of characteristic parameters, the adjustments of resources, and the reward values; and

constructing the action-reward fuzzy rules based on the membership functions.

13. The method according to claim 1, wherein the deviated reward value is given higher if a certain amount of at least one resource is released from the workload processed in the step G.

14. The method according to claim 1, wherein a gamma value times to each reward values before accumulating with the deviated reward value in the step J.

15. The method according to claim 14, wherein the gamma value ranges from 0 to 1.