Hi Robert,

Were you running MPI_Gather in multiple threads concurrently on a same communicator?

That is not allowed. You’ll need at least different communicator for different threads.

If that was not the issue, could you try compile MPICH with ch4, with `--with-device=ch4:ofi`, assuming you are using the latest release.

-- 
Hui Zhou

From: Robert Katona <robert.katona@hotmail.com>
Date: Wednesday, January 13, 2021 at 10:34 AM
To: Zhou, Hui <zhouh@anl.gov>, devel@mpich.org <devel@mpich.org>
Cc: Brent Morgan <brent.taylormorgan@gmail.com>
Subject: Re: [mpich-devel] mpich3 error

Hello Hui Zhou,

And can you help us in brainstorming?

Unfortunately we cannot share the code. But I can give you more details. 

Couple of computers are connected into the same network and we would like to run a distributed calculation. During testing we run the MPI application over 50 job executer processes/threads and one job evaluter master/root thread. Each job calculate two float value what the master collects. To distribute this data we use the MPI_Gather() function. The execution of the problem runs fine but randomly stops with error message what Brent sent to you.

Sometimes it stops after 10 iterative calculation, sometimes after 50 or 70. We do the same calculation, with same input, but the event of the error is very random. And it always stops in the MPI_Gather() function, when the master try to collect the data from all of the jobs.

If there is a network issue with any of the computer can it produce this error?

We are using Mpich 3.3.2, in the release note of 3.4 it is written the network communication has changed from ch3 to ch4. With ch4 can we expect better behavior?

Can you give us any hint where to look and what to check in this topic?

Regards,

Robert

Az Android Outlook letöltése

---

Feladó: Zhou, Hui <zhouh@anl.gov>
Elküldve: 2021. január 13., szerda 15:51
Címzett: devel@mpich.org
Másolat: Brent Morgan; Robert Katona
Tárgy: Re: [mpich-devel] mpich3 error

Hi Brent,

Unfortunately, unless you can provide us with a reproducer, there is little we can do to find the issue.

-- 
Hui Zhou

From: Brent Morgan via devel <devel@mpich.org>
Date: Wednesday, January 13, 2021 at 3:20 AM
To: devel@mpich.org <devel@mpich.org>
Cc: Brent Morgan <brent.taylormorgan@gmail.com>, Robert Katona <robert.katona@hotmail.com>
Subject: [mpich-devel] mpich3 error

Hello mpich dev support,

I am receiving the following error in my MPI implementation, when I use 110+ threads.

Assertion failed in file src/mpid/ch3/channels/nemesis/src/ch3_progress.c at line 567: !vc_ch->recv_active
0x7f7f870937 ???
???:0
0x7f7f870957 ???
???:0
0x7f7f8089cf ???
???:0
0x7f7f854983 ???
???:0
0x7f7f75a4f3 ???
???:0
0x7f7f7cd8bf ???
???:0
0x7f7f7cde0f ???
???:0
0x7f7f78d577 ???
???:0
0x7f7f6ea6cb ???
???:0
0x7f7f6ea8f3 ???
???:0
0x7f7f78c7ff ???
???:0
0x7f7f6e9e6f ???
???:0
0x7f7f6e9eb7 ???
???:0
0x7f7f6ea033 ???
???:0
0x55793fe47b ???
???:0
0x55793f96b3 ???
???:0
0x7f7f29908f ???
???:0
0x55793f9a9b ???
???:0
internal ABORT - process 171

Does this give a clue as to what I may be doing wrong?  Thank you,

Best,

Brent