[mpich-discuss] FAQ about multicore clusters

[Rajeev Thakur]

For multicore clusters, you should use the Nemesis channel in MPICH2. In
1.1a1 it is the default; in 1.0.7 you need to configure with
--with-device=ch3:nemesis. Nemesis will use shared memory for communication
within a node and TCP (or other network) across nodes. For how to configure
and run MPD, see Section 5.1.7 "Running MPD on SMPs" of the MPICH2
installation guide.

Rajeev 

> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Nicholas Yue
> Sent: Thursday, September 25, 2008 7:58 AM
> To: mpich-discuss at mcs.anl.gov
> Subject: [mpich-discuss] FAQ about multicore clusters
> 
> Hi,
> 
>     I wasn't able to search the various mailing list archives so I am 
> posting here in the hope that someone can point me to more 
> documentation 
> or information.
> 
>     I have used MPICH on and off since version MPICH1 days.
> 
>     I have recently got hold of a dual quad core Xeon and 
> have build the 
> multi-threaded version of MPICH2.
> 
>     In the coming months, I am involved in the deployment of a 
> renderfarm (i.e. Gigabit network, not Myrinet type of connectivity) 
> which I'd like to run some MPI feasibility tests.
> 
>     Is there some documentation or information about such hybrid i.e. 
> multi-core cluster, specifically, what configure options to use for 
> building the source? What would be the way to configure and run mpd? 
> e.g. ncpus, etc
> 
>     As a hypothetical case,  let's just say
> 
>     * 10 Gigabit  TCP/IP interconnect
>     * Dual Quad Core Xeon per machine/1U
>     * 8 Gig RAM
>     * 64 machines
> 
> 
> Regards
> 
>