MSCCL++ v0.3.0 Release Plan

[chhwang]

Timeline

Released Date: Oct. 11th, 2023

Work Items

Communication Features and Interfaces

GPU-side communication interfaces

1. • [ ] Interfaces: add tile-based communication interfaces (Add tile-based API #145) moved to the next version
2. • Semaphore: add timeout for semaphores ([Feature] semaphores do not take a timeout for wait. #161)
3. • Interfaces: align interfaces of put(), get(), putPackets(), and getPackets() ([Feature] getPacket arg list does not match the get function from sm_channels #158)

Host-side interfaces

1. • Python: add Python bindings (Python bindings #125)
2. • Python: [feature] add .def_prop_ro for all device components so that they can be passed to GPU kernels. #144
3. • [ ] SmChannel: [feature] sm_channel constructors need to be improved. #148 moved to the next version
4. • Configuration: [feature] Enable custom fifo size and tail flush #138, [feature] all of #define stuff need to be an arg #149, Extend configurations #150
5. • Interfaces: connection setup without bootstrap (Add Context and Endpoint classes to enable non-Communicator use-cases #166)
6. • Interfaces: construct ProxyChannel with shared pointers (Construct ProxyChannel with shared pointers #184)

Bug Fix

1. • Bug: semaphore ownership issues (Make sure the semaphore not be released during the lifecycle of SmChannel #131)
2. • Bug: [bug] device functions really need to be in a different file #139, Change device handle interfaces & others #142
3. • Bug: [bug] change all __CUDACC__ to __CUDA_ARCH__ #143
4. • Bug: [Bug] Bootstrap occasionally returns "Address in use" error during initialize() #163
5. • Bug: compilation fails in the debug mode ([Bug] Compilation fails with CMAKE_BUILD_TYPE=Debug #174)

Performance Optimization

1. • Latency: [Performance] Improve single-node AllReduce latency #164

Development Pipeline

1. • Unit Tests: add Python unit tests ([Unit Test] Add Python unit tests #134)
2. • Unit Tests: add more unit tests ([Unit Test] Listing unit tests needed #115)
3. • [ ] Debugging Tool: fix broken NPKit support (@yzygitzh)
4. • Debugging Tool: support POLL_PRINT_ON_STUCK (Change device handle interfaces & others #142)

Documents

1. • [ ] Docs: update performance numbers on readme (call for contribution) moved to the next version
2. • Docs: Update install guidelines #159

[chhwang]

Released: https://github.com/microsoft/mscclpp/releases/tag/v0.3.0