[MPICH] MPD caused major system throughput degration, Redhat ES 4, U2

[Ralph Butler]

Are you sure you have the latest release of mpich2.  I ask this
because a long time ago there was a bug in mpd that caused excessive
checking of sockets in the main loop.

I logged in to each of 3 machines running:
     Debian Linux
     Fedora Core 5
     Mac OS X
I was the only user on each machine and each initially had:
     load average: 0.00, 0.00, 0.00

On every host I ran the same job (that took at least 124 seconds
depending on the platform).  I ran the job several times, both with
and without MPD running.  The time never varied by no more than
0.1 second on a single platform.

--ralph

On ThuAug 17, at Thu Aug 17 4:13PM, chong tan wrote:

> on SUN Ultra 40, 2 Opterons 280, 8G memory.
> The application does not use mpi at all
>
> time application .....
> 627.007u ....... 99.9% ....
>
> if I start mpd before the application:
>
> mpd &
> time application .....
> 860.365u ........ 99.9% ...
>
> If I kill the mpd using mpdallexit, and I rerun the
> application, I get almos the same time.
>
> The application is the only one running on the box,
> cause I am still in the process of setting up that
> box.
>
> any suggestion ?
>
> tan
>
>
> __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com
>
>