Changelog

Note

Code is maintained in https://github.com/fengggli/gpu-computing-materials/

Ask Feng for access.

Current

Date:2019-11-25

Added

  1. Flexflow paper(https://fengggli.github.io/ResearchDocs/topics/hybridparal/index.html#beyond-data-and-model-parallelism-for-deep-neural-networks) from Alex Aiken’s group(whose previous work includes Sequoia and Legion), architecture topology is considered for GPU clusters.
  2. Channel and Filter parallelism (https://fengggli.github.io/ResearchDocs/journal/Fall19/Week13.html#channel-and-filter-parallelism)

Working on

  1. Reorganize code thread model, so that different layers could have different parallelism.
  2. Performance evaluation on different model/data paralellism setup.

TODO List

  • Use fine-grained lock to reduce contention.
  • Theoretical model.