I think TPE stands for thread progress engine and the question is about the asynchronous progress agent enabled by MPICH_ASYNC_PROGRESS=1. This agent is very high level and I doubt there’s a good way to make it support blocking mode. You’d be better off writing a new async progress engine in the device code. Another idea is to not touch MPICH and instead look at OFI/libfabric, which supports different progress modes and may be more amenable to what you want. You should be able to get pointers from Sean Hefty and friends on the associated email list. Jeff On Fri, Oct 20, 2017 at 4:26 PM Halim Amer <[email protected]> wrote:

Perhaps you could tell us a bit more about

1) what you mean by "progress engine to be enabled". Any device implementation in MPICH has a progress engine enabled all the time. There is no such thing as enable/disable the progress engine. Removing the progress engine would simply cause processes to hang

2) what TPE stands for

Halim www.mcs.anl.gov/~aamer

On 10/20/17 12:12 PM, Pau Farré wrote:

...

Hi

Here at BSC (Antonio Peña's group) we were doing some modifications to mpich that require the progress engine to be enabled, but we found that the TPE polling was introducing some overhead.

We were wondering if mpich supports a 'blocking' mode for the TPE. And if so, It is compatible with Intel Omni-Path?

Thanks. Pau Farré

http://bsc.es/disclaimer _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

---

discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

-- Jeff Hammond [email protected] http://jeffhammond.github.io/ _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss