Re: MPI debugging workflows

[Alastair McKinstry <mckinstry@debian.org>]

On 04/08/2018 22:41, Dima Kogan wrote:
> Answering some of my own questions inline...
>
>
> Dima Kogan <dima@secretsauce.net> writes:
>
> > 2. Is the MPI implementation significant? Would mpich behave
> > potentially differently here from openmpi?
> I rebuilt sundials with mpich instead of openmpi, and those tests now
> work just fine: nothing locks up. There might be other issues involved
> here, but I'd like to fully figure out what's going on before proposing
> any such change.
>
> I also poked around with a debugger a little bit: the lockup is inside
> some MPI calls. Anybody have experience here? I can imagine this is an
> openmpi bug, but the last time I touched MPI anything was about 20 years
> ago, so any info from more-recently-experienced people would be welcome.
There have been a number of lockup issues with openmpi recently, 
including thread lockups in underlying pmix libs.

There is a new 3.1.2rc2 release that I'm testing. Can you give a 
reproducible test case for me to test against, as I think this release 
may have the necessary fix.

regards
Alastair

--
Alastair McKinstry, <alastair@sceal.ie>, <mckinstry@debian.org>, https://diaspora.sceal.ie/u/amckinstry
Misentropy: doubting that the Universe is becoming more disordered.