[GE users] Comparison of MPI Job Runtimes with and without SGE

Reuti reuti at staff.uni-marburg.de
Wed Aug 11 22:44:35 BST 2004

    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]


>    I was running an MPI job on my solaris cluster. I found out that the 
>job actually takes less amount of time to execute with MPICH tightly 
>integrated with SGE than when it runs through plain MPICH only. Note 
>that there is no other user load on the cluster. and this MPI job is 
>only user job running on the cluster. and most impoprtantly i am feeding 
>  eaxctly the same set of machines to the job (when it running through 
>plain MPICH) as SGE_tight integration makes the job run on.
>I would assume, given that both times job is run on the same set of 
>nodes, SGE should incur more cost as comapred to MPICH?. Any idea ?

what MPICH device and version are you using: ch_p4, Myrinet, ... mpich ... ssh/rsh

Interesting, I never thought about this. I depends on the amount of "rsh" calls 
I could imagine. If you have a parallel job, where all the nodes are getting 
only one rsh-command and calculating for hours, then there shouldn't be a 
difference. When you have many small steps with many rsh-calls, a rsh-daemon 
has to be started by SGE with each qrsh. Maybe this is faster than the login in 
to the default daemon (i.e. going through inetd.conf ... starting session)

With the ch_p4 device, there is the possibility to speed up the creation of the 
sessions by starting a server on the nodes (serv_p4). Can you try this (without 

I'm curious, whether there is also a difference.

Cheers - Reuti

To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net

More information about the gridengine-users mailing list