when to use flagcxBarrier?

[singer007]

hi,

I have seen some code in this file
https://github.com/FlagOpen/FlagCX/blob/main/test/perf/test_sendrecv.cpp#70

        for(int i=0;i<num_iters;i++){
            flagcxGroupStart();
            flagcxSend(sendbuff, count, DATATYPE, sendPeer, comm, stream);
            flagcxRecv(recvbuff, count, DATATYPE, recvPeer, comm, stream);
            flagcxGroupEnd();
        }
        flagcxBarrier(comm, stream);

Is the flagcxBarrier here works the same as devHandle->streamSynchronize(steam) ?

why here use flagcxBarrier, It add more time when timing the performance of send/recv.

and what does flagcxBarrier used for ?

[MC952-arch]

Yep, you may use devHandle->streamSynchronize(stream) to achieve synchronization. Here flagcxBarrier is added to testify barrier op in all test cases.

Since flagcxBarrier introduces an extra allreduce operation, it will cause performance degradation for sure.

Thank you for your feedback. We will fix this issue as soon as possible.

[singer007]

Hi, may I submit a PR to fix this issue?

[MC952-arch]

Hi, may I submit a PR to fix this issue?

Hi, this issue has already been fixed in the latest version.
If you encounter any other issues while using it, feel free to submit an issue and a PR.

[singer007]

ok, thanks