Skip to content

MSCCL++ v0.2.0

Compare
Choose a tag to compare
@chhwang chhwang released this 11 Jul 03:03
2e16457

Communication Features and Interfaces

GPU-side communication interfaces (DeviceChannel)

    • Proxy-based Interfaces: ProxyChannel (#66)
    • In-SM Copy Interfaces: SmChannel (#55)
    • Packet Copy Interfaces: putPackets(), getPackets(), signalPacket() (#85, #90, #102)

Host-side interfaces

    • Bootstrap: fix socket performance issue & bugs (#92, #100, #113)
    • Communicator: implement (#66)

Transports support

    • NVLink: implement (#66)
    • InfiniBand: implement (#66)
    • InfiniBand: tackle memory consistency issues (#96)

Performance Optimization

    • Throughput: pass AllGather perf qualification (#77)
    • Throughput: pass AllReduce perf qualification (#83, #90)
    • Throughput: pass AllToAll perf qualification (#87)
    • Latency: pass AllReduce perf qualification (#85, #90)
    • Latency: pass 2-node AllReduce perf qualification (#109, #118)

Development Pipeline

    • Unit Tests: cover all interfaces (#81, #91)
    • mscclpp-test: add AllGather (#77)
    • mscclpp-test: add AllReduce (#83)
    • mscclpp-test: add AllToAll (#87)
    • CI: lint, spelling, CodeQL (#79)
    • CI: unit test (#81)
    • Package: publish Docker images (#104)

Documents

    • Doxygen: add configuration (#72)
    • README: enhance details (#88)
    • License: add license comments on all files (#106)
    • Code: cleanup & comments (#86, #119)

Full Changelog: https://github.com/microsoft/mscclpp/commits/v0.2.0