Hui Zhou comments

Results 695 comments of


                                            Hui Zhou

osu_gather and osu_gather_persistent fails starting 8 nodes 96 ppn

We found the bug: * MPI_Gather with sufficient message size and sufficient number of processes, it goes to the gather_intra_binomial algorithm, which does message combination. The algorithm constructs a `struct`...

Have valgrind suppression files installed

Thanks for creating the issue. MPICH itself do not have valgrind suppression file - we may used to have, but I don't see any now. The ones you found is...

Have valgrind suppression files installed

Depend on what noise you are getting with Valgrind, for example, with ucx, I think you can grab the one from ucx distribution, If you don't really care about which...

Have valgrind suppression files installed

We have not tested with the shm provider extensively, so I don't know. The proof is in the pudding. We currently use `-fsanitize=address` in our CI testing.

osu_igather Other MPI error starting at 8 nodes 96 ppn

@longfei-austin 8 nodes at what PPN?

osu_igather Other MPI error starting at 8 nodes 96 ppn

I am using 8 nodes and the `aurora_test` branch, running `osu_igather` - it works for me all the way up to 96 ppn.

[MPIX Stream] CUDA event compatible completion semantics

I agree the event should observe the enqueued operations. I need examine the implementation. Do you have a reproducer?

osu_ialltoall Other MPI error starting at 32 nodes 96 ppn

Are you running at 96 PPN? I suspect the error is a libfabric I/O error, similar to the other issues when the network is overwhelmed and libcxi bails out.

osu_gather significant slowdown from 8b to 1024b

Please show the complete benchmark results from 1 byte to 4096 bytes

MPI-IO (ROMIO) collective buffering performance bottleneck on Aurora

Is this results considered acceptable ? ``` write 175063.57 119909.89 155586.28 19059.75 2735.37 1873.59 2431.04 297.81 6.85890 NA NA 0 16384 64 5 0 1 1 0 0 1 67108864...