Re: [Myrinet] Using mpd with gm apps in mpich-1.2.4..8a (performance difference)

Hi Craig,

On Tue, 2002-09-10 at 17:30, Craig Tierney wrote: 
> I am trying to configure my cluster to run
> using the MPD process to help speedup large
> The bandwidth is about 1/2 of what I expect and
> the latency is about 50% worse when using mpd to
> start the program.  Am I doing something wrong?

I get identical results with or without MPD. It's as expected because
MPD is involved only at startup time. 

> #p0     p1      dist    len     ave time (us)   rate
> 0       1       1       0       14.569375       0.00
> 0       1       1       1048576 7687.510000     136.400e+6

It looks like loopback performance instead of point-to-point. Are you
sure that the 2 processes are spawned on 2 different nodes ?
BTW, "-y" is specific to the Myrinet usage at ANL, it's useless for
MPICH-GM.

You can do 2 things: 
* check that the processes are effectively spawned on 2 different nodes
(print the host name in the mpptest code). 
* check that no other previous mpptest instances are still running on
the same nodes. 

I have tried to reproduce your case in many many ways, without success.

Patrick 
-- 
----------------------------------------------------------
|   Patrick Geoffray, Ph.D.      [email protected] 
|   Myricom, Inc.                http://www.myri.com
|   Cell:  865-389-8852          685 Emory Valley Rd (B)
|   Phone: 626-821-5555          Oak Ridge, TN 37830
----------------------------------------------------------