[mpich-discuss] Hydra issues

[Scott Atchley]

On Aug 27, 2009, at 8:43 PM, Scott Atchley wrote:

> On Aug 27, 2009, at 5:43 PM, Pavan Balaji wrote:
>
>>>> The first one is an MPI application, so the output should get  
>>>> printed as it's generated. The second one is not using MPI, so  
>>>> the output should get buffered till the end.
>>> Yes, it works as you describe.
>>
>> Ok, now I'm totally confused. I just tried IMB-3.0 and the output  
>> gets printed as it is generated. So, what are we doing differently?  
>> I tried 1.1.1p1 and that worked fine as well.
>
> Let me start over tomorrow with clean installs of MPICH2 and IMB. I  
> will try it in-house and at another site.
>
> Scott

Using two machines in-house, I downloaded both MPICH2 1.1.1.p1 and IMB  
3.2. I compiled a default MPICH2:

$ ./configure --with-device=ch3:nemesis --prefix=$PWD/gcc

Note, I am using TCP here and I also tries ch3:nemesis:mx with the  
same results.

I then compiled IMB using this MPICH2. I then ran IMB using Hydra:

$ cat hosts
wind01:2
wind04:2

$ mpiexec.hydra -f hosts -n 4 ./IMB-MPI1

and I see _no_ output until IMB completes.

If I boot the proxies first and then run IMB, output starts immediately.

I tried your test app again and it starts immediately without NO_MPI  
set and is delayed when NO-MPI=1; the exact opposite of the above. I  
tested IMB 2.3 and it too is delayed.

Scott