[mvapich-discuss] Questions in OSU OpenSHMEM Microbenchmarks

[Hashmi, Jahanzeb]

Hi Naveen,


Your observation about the shmem_malloc/shmalloc call using the number of elements instead of bytes is correct. We appreciate you taking time to report the issue. We have fixed applied the appropriate fix to the osu_oshm_reduce benchmark and the changes will be available in the next OMB release.


Feel free to let us know if you encounter any other issue.


regards

Jahanzeb

________________________________
From: Naveen Ravichandrasekaran <nravi.research at gmail.com>
Sent: Tuesday, December 19, 2017 4:29:17 PM
To: Hashmi, Jahanzeb
Cc: mvapich-discuss at cse.ohio-state.edu
Subject: Re: [mvapich-discuss] Questions in OSU OpenSHMEM Microbenchmarks

Hi Jahanzeb,
Thanks for reply.
You are correct about the spec. The spec speaks about the size of pWrk in terms of the number of elements.
In the benchmark, MAX(nreduce/2+1, _SHMEM_REDUCE_MIN_WRKDATA_SIZE) gives you the correct number of elements.

But, the issue here is the symmetric heap allocation. As per the spec, when we allocate the heap using shmem_malloc/shmalloc routines - we need to pass the size of the array in bytes and not just the number of elements.
So, it still looks as an issue to me at that shmem_malloc/shmalloc operation. Can you please verify?





On Tue, Dec 19, 2017 at 2:36 PM, Hashmi, Jahanzeb <hashmi.29 at buckeyemail.osu.edu<mailto:hashmi.29 at buckeyemail.osu.edu>> wrote:

Hi Naveen,

The OpenSHMEM (1.3) specification document (page 58) states that:
"The pWrk argument must have the same data type as dest. In C/C++, this contains
max(nreduce/2 + 1, SHMEM_REDUCE_MIN_WRKDATA_SIZE) elements."

http://www.openshmem.org/site/sites/default/site_files/OpenSHMEM-1.3.pdf

The osu_oshm_reduce benchmark in OMB is following the OpenSHMEM standard. Further, the collective benchmarks in OpenSHMEM report number of elements instead of bytes as max_msg_size is 4M that corresponds to 1M float elements.

Please let us know if there is any other issue.

Thanks

Jahanzeb

_______________________________________________
mvapich-discuss mailing list
mvapich-discuss at cse.ohio-state.edu<mailto:mvapich-discuss at cse.ohio-state.edu>
http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20171219/69eae30f/attachment.html>