COORDINATED VERSION CONTROL SYSTEM, METHOD, AND RECORDING MEDIUM FOR PARAMETER SENSITIVE APPLICATIONS

• US 20170344904A1
• Filed: 05/31/2016
• Published: 11/30/2017
• Est. Priority Date: 05/31/2016
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Hack, Michel H.T., Ren, Yufei, Wang, Yandong, Zhang, Li
• Granted Patent: US 10,657,459 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/1873 Versioning file systems, temporal file systems, e.g. file system supporting different historic versions of files
• CPC Class Codes: G06F 16/219 Managing data history or versioning querying versioned data G06F16/2474; querying temporal data G06F16/2477
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Version vector-based rules are used to facilitate asynchronous execution of machine learning algorithms. The method uses version vector based rule to generate aggregated parameters and determine when to return the parameters. The method also includes coordinating the versions of aggregated parameter sets among all the parameter servers. This allows to broadcast to enforce the version consistency; generate parameter sets in an on-demand manner to facilitate version control. Furthermore the method includes enhancing the version consistency at the learner'"'"'s side and resolving the inconsistent version when mismatching versions are detected.

Claims

1. A coordinated version control system including a parameter server having a plurality of parameter sets and a plurality of learners communicating with the plurality of parameter sets, the system comprising:

a leader parameter collecting circuit configured to collect parameter set data for a leader parameter set from the plurality of learners;

a follower parameter collecting circuit configured to collect parameter set data for a follower parameter set from the plurality of learners;

a leader aggregated parameter generating circuit configured to generate a new leader version of a leader aggregated parameter set based on the leader parameter set data;

an event broadcasting circuit configured to generate a broadcast event indicating that the new leader version of the leader aggregated parameter set has been generated;

a broadcast detecting circuit configured to detect the broadcast event;

a checking circuit configured to check if a version of the follower parameter set matches the new leader version of the leader aggregated parameter set based on the broadcast detecting circuit detecting the broadcast event; and

a follower aggregated parameter generating circuit configured to generate a new follower version of a follower aggregated parameter that matches the new leader version of the leader aggregated parameter set.

2. The system of claim 1, wherein the follower aggregated parameter generating circuit only generates the new follower version of the follower aggregated parameter if the checking circuit does not determine the version of the follower parameter set matches the leader new version of the leader aggregated parameter set.

3. The system of claim 1, further comprising:

a leader pull requesting circuit configured to authorize a request from the plurality of learners to pull the new leader version of the leader aggregated parameter set; and

a follower pull requesting circuit configured to authorize a request from the plurality of learners to pull the new follower version of the follower aggregated parameter set.

4. The system of claim 3, wherein the leader pull requesting circuit immediately authorizes the request when the leader aggregated parameter generating circuit generates the new leader version of the leader aggregated parameter set.

5. The system of claim 3, wherein the follower pull requesting circuit authorizes the request when the follower aggregated parameter generating circuit generates the new follower version of the follower aggregated parameter that matches the new leader version of the leader aggregated parameter set.

6. The system of claim 1, wherein the leader aggregated parameter generating circuit includes:

a server checking circuit configured to check if the leader aggregated parameter generating circuit should generate the new leader version of the leader aggregated parameter set based on an intermediate version vector for each of the plurality of learners;

an epoch update checking circuit configured to check a number of updated epochs based on a number of the plurality of learners that have pushed updated parameter set data to the parameter server according to the intermediate version vector; and

an aggregation triggering circuit configured to cause the leader aggregated parameter generating circuit to generate the new leader version of the leader aggregated parameter set if the number of updated epochs is equal to the number of the plurality of learners that have pushed the updated parameter set data to the parameter server multiplied by a predetermined threshold portion of the plurality of learners.

7. The system of claim 1, wherein the leader parameter set data is collected by the leader parameter collecting circuit for a previous epoch.

8. The system of claim 1, wherein the leader aggregated parameter generating circuit generates the new leader version of the leader aggregated parameter set for each of the plurality of learners to pull at a current epoch.

9. The system of claim 7, wherein the leader aggregated parameter generating circuit generates the new leader version of the leader aggregated parameter set for each of the plurality of learners to pull at a current epoch.

10. The system of claim 1, wherein the leader is selected by any one of:

a Raft leader election protocol;

a Zab leader election protocol; and

a Paxos leader election protocol

11. A coordinated version control method for a parameter server having a plurality of parameter sets and a plurality of learners communicating with the plurality of parameter sets, the method comprising:

first collecting parameter set data for a leader parameter set from the plurality of learners;

second collecting parameter set data for a follower parameter set from the plurality of learners;

generating a new leader version of a leader aggregated parameter set based on the leader parameter set data;

broadcasting a broadcast event indicating that the new leader version of the leader aggregated parameter set has been generated;

checking if a version of the follower parameter set matches the new leader version of the aggregated parameter set based on the broadcast event being detected; and

generating a new follower version of a follower aggregated parameter that matches the new leader version of the leader aggregated parameter set.

12. The method of claim 11, wherein the generating the new follower version only generates the new follower version of the follower aggregated parameter if the checking does not determine the version of the follower parameter set matches the leader new version of the leader aggregated parameter set.

13. The method of claim 11, further comprising:

first requesting authorization of a request from the plurality of learners to pull the new leader version of the leader aggregated parameter set; and

second requesting authorization of a request from the plurality of learners to pull the new follower version of the follower aggregated parameter set.

14. The system of claim 13, wherein the first requesting authorization immediately authorizes the request when the generating the new leader version generates the new leader version of the leader aggregated parameter set.

15. The method of claim 13, wherein the second requesting authorization authorizes the request when the generating the new follower version generates the new follower version of the follower aggregated parameter that matches the new leader version of the leader aggregated parameter set.

16. The method of claim 11, wherein the generating the new leader version further comprises:

checking if the generating the new leader version should generate the new leader version of the leader aggregated parameter set based on an intermediate version vector for each of the plurality of learners;

checking a number of updated epochs based on a number of the plurality of learners that have pushed updated parameter set data to the parameter server according to the intermediate version vector; and

triggering the generating the new leader version to generate the new leader version of the leader aggregated parameter set if the number of updated epochs is equal to the number of the plurality of learners that have pushed the updated parameter set data to the parameter server multiplied by a predetermined threshold portion of the plurality of learners.

17. The method of claim 11, wherein the leader parameter set data is collected by the first collecting for a previous epoch.

18. The method of claim 11, wherein the generating the new leader version generates the new leader version of the leader aggregated parameter set for each of the plurality of learners to pull at a current epoch.

19. A non-transitory computer-readable recording medium recording for coordinated version control performing the method of claim 11.

20. An asynchronous execution facilitation system including a parameter server having a plurality of parameter sets and a plurality of learners communicating with the plurality of parameter sets, the system comprising:

a server checking circuit configured to check if a new version of an aggregated parameter set can be generated based on an intermediate version vector for each of the plurality of learners;

an epoch update checking circuit configured to check a number of updated epochs based on a number of the plurality of learners that have pushed updated parameter set data to the parameter server according to the intermediate version vector; and

an aggregation triggering circuit configured to generate the new version of the aggregated parameter set if the number of updated epochs is equal to the number of the plurality of learners that have pushed the updated parameter set data to the parameter server multiplied by a predetermined threshold portion of the plurality of learners.

21-30. -30. (canceled)