Job launch issue with Intel MPI

[budana...@gmail.com]

Hi,

Since a recent update, job launch with Intel-MPI fails for Intel-MPI starting with 2019. Job launch works fine with OpenMPI. The LRZ support advised me to not use OpenMPI right now as in our supercomputer, it might have some inconsistency and latency issues. 

But since I compile with UPCXX_NETWORK=ibv, the MPI library is actually only needed for job launch so can I conclude that having a slow MPI implementation has no possibility to make my UPC++ program slower, with the sole exception of needing more time for the job launch. Am I right about that?

[Paul H. Hargrove]

Yes, you are correct that with UPCXX_NETWORK=ibv the MPI implementation is used only to launch the job.

So its performance is not important other than the time spent to startup and shutdown the job.

-Paul

--
You received this message because you are subscribed to the Google Groups "UPC++" group.
To unsubscribe from this group and stop receiving emails from it, send an email to upcxx+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/upcxx/ef45b33e-b780-460b-8691-752b729a5eaen%40googlegroups.com.

--

Paul H. Hargrove <PHHar...@lbl.gov>
Pronouns: he, him, his

Computer Languages & Systems Software (CLaSS) Group

Computer Science Department

Lawrence Berkeley National Laboratory

[Dan Bonachea]

On Fri, May 20, 2022 at 7:00 PM Paul H. Hargrove <phhar...@lbl.gov> wrote:

Yes, you are correct that with UPCXX_NETWORK=ibv the MPI implementation is used only to launch the job.

So its performance is not important other than the time spent to startup and shutdown the job.

An important caveat to this answer:

Using MPI for spawning means the MPI library is initialized into the process and still consumes some memory and adaptor resources for the life of the job, even if there is no active MPI communication happening after job spawn. This is usually not something to worry about, unless your application is very memory-hungry or running at very large scales.

-D

 

-Paul

On Fri, May 20, 2022 at 7:14 AM budana...@gmail.com <budana...@gmail.com> wrote:

Hi,

Since a recent update, job launch with Intel-MPI fails for Intel-MPI starting with 2019. Job launch works fine with OpenMPI. The LRZ support advised me to not use OpenMPI right now as in our supercomputer, it might have some inconsistency and latency issues. 

But since I compile with UPCXX_NETWORK=ibv, the MPI library is actually only needed for job launch so can I conclude that having a slow MPI implementation has no possibility to make my UPC++ program slower, with the sole exception of needing more time for the job launch. Am I right about that?

--
You received this message because you are subscribed to the Google Groups "UPC++" group.
To unsubscribe from this group and stop receiving emails from it, send an email to upcxx+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/upcxx/ef45b33e-b780-460b-8691-752b729a5eaen%40googlegroups.com.

--

Paul H. Hargrove <PHHar...@lbl.gov>
Pronouns: he, him, his

Computer Languages & Systems Software (CLaSS) Group

Computer Science Department

Lawrence Berkeley National Laboratory

--
You received this message because you are subscribed to the Google Groups "UPC++" group.
To unsubscribe from this group and stop receiving emails from it, send an email to upcxx+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/upcxx/CAAvDA162Z5Ljwemu8a1PO50Z0srrhHp%3Dt1oQ8U-AGRVKJrSpZQ%40mail.gmail.com.