Details
-
Story
-
Status: To Do
-
Major
-
Resolution: Unresolved
-
None
Description
Currently, multithreading in training is achieved through OpenMP at operator level. There are issues filed on GitHub about OpenMP in MXNet. Besides, there is no clear benchmark on the performance impact on various platforms
Acceptance Criteria
- Fix/address all issues related to OpenMP
- Understand the performance impact of OpenMP with benchmarks
- Provide an optimal thread setting for users at different computing platforms
Attachments
Issue Links
- requires
-
MXNET-1051 Fix the current issue in OpenMP
- To Do
-
MXNET-1053 Performance benchmark with and without multithreaded training
- To Do
-
MXNET-1056 Implement an algorithm to select the optimal number of threads in OpenMP for each platform
- To Do