Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2006-07-20 10:41:11

I think there are two questions here:

1. Running MPI applications on "slow" networks (e.g., 100mbps). This is
very much application-dependent. If your MPI app doesn't communication with
other processes much, then it probably won't matter. If you have
latency/bandwidth-sensitive applications, then using a "slow" network can
definitely have a negative impact on performance.

2. Running MPI applications on resources that are being used by others. In
this case, your MPI processes will be competing with other processes for
CPU, RAM, and other resources -- just like any other process. Hence, your
overall performance will depend not only on the application, but also on the
usage patterns of the other resources (e.g., the workstations and the people
that use them).

I have certainly heard of bunches of success stories in this kind of
environment -- small numbers of relatively lightly-loaded workstations
(typically <= 16) running small to mid-sized MPI applications, etc. A
common case for such scenarios is for development and debugging, or even
running small versions of jobs when you can't get time on larger resources,
etc. Specifically: sometimes running a smaller version of your job is
better than not running anything at all.

Hope that helps.

On 7/20/06 10:04 AM, "Brock Palen" <brockp_at_[hidden]> wrote:

> Its doable, the scaling will not as good, because a network is a
> network. If you are using just regular 100Mbit, you will not scale
> as far as really good 1gig ethernet, but we are still talking about
> tcp which incurs a penalty over networks like infiniband and myrinet.
> Tcp is the largest issue, its going to be really application
> dependent you are right.
> On another note though many of the older cluster that are now out of
> service used just 100Mbit ethernet and worked.
> Brock Palen
> Center for Advanced Computing
> brockp_at_[hidden]
> (734)936-1985
> On Jul 20, 2006, at 9:27 AM, Vladimir Sipos wrote:
>> Hi,
>> Is MPI paradigm applicable to the cluster of regular networked
>> machines.
>> That is, does the cost of network IO offset benefits of
>> parallelization?
>> My guess is that this really depends on the application itself,
>> however,
>> I'm wondering if you guys know of any success stories which involve
>> MPI
>> running on a set of networked machines (not beowulf cluster or any
>> SC).
>> Thanks,
>> Vladimir Sipos
>> Software Engineer
>> Advertising Technology
>> CNET Networks, Inc.
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
> _______________________________________________
> users mailing list
> users_at_[hidden]

Jeff Squyres
Server Virtualization Business Unit
Cisco Systems