I was going to start small given the budget I have. Eventually, I'd
like dedicate a gigabit switch for HTTP traffic and Infiniband for
compute traffic. At first, I don't expect too much MPI work to be
done, but I've heard FreeBSD performing better under duress than linux
as the number of HTTP threads increases.
Knowing that beowulf stuff is done better on linux another option
would be to run FreeBSD inside of Xen for HTTP, while Linux does
computing.
How good is FreeBSD for clustering? I'm not really familiar with
FreeBSD for that use so much and there isn't a lot of documentation
for FreeBSD for building beowulfs.
The final option would be to divide and conquer: 6 for HTTP, 6 for
computing, but my reasoning is why not scale for HTTP as much as
possible.
In this setup, HTTP would be primary deal, which was why I went to
FreeBSD first. Does OpenMPI or MPICH2 run well under FreeBSD? I got a
build working on OpenBSD/sparc64, but haven't really done much with it
yet.
Thanks for the help,
Vivek
On Thu, May 7, 2009 at 9:55 AM, Will Maier <willma...@ml1.net> wrote:
> Hi Vivek-
>
> On Thu, May 07, 2009 at 09:36:17AM -0700, Vivek Ayer wrote:
>> 1 OpenBSD Router running 4.5 routing to a subnet of 13 nodes running
>> FreeBSD 7.2. Of the 13 nodes, 1 node is a master mysql server and the
>> 12 nodes will run apache running LAMP-like services. The router will
>> round-robin using hoststated for load-balancing.
>
> There are some FreeBSD clusters out there (NCSA has one, IIRC), but
> they're certainly not as common as Linux. If your users can run on
> FreeBSD, you might as well use it. If their code is all Linuxy (and lots
> of cluster and -- even more so -- grid code make silly assumptions like
> that), you should give them a platform that they can easily use.
>
>> However, they will serve an additional task: The master mysql server
>> will be head node for MPI jobs delivered to the 12 nodes. Basically,
>> this setup will double up as a beowulf and web server. Is this
>> efficient? I imagine the MPI jobs won't be running all the time and
>> while they're up, might as well do something.
>
> This might work. But you're setting yourself up for contention and
> degraded service to at least one set of users. Do the people who care
> about perfomance of your LAMP stack mind waiting a bit while MPI jobs
> chew memory and network bandwidth? Do your MPI users mind if their jobs
> take longer to complete while your LAMP stuff is getting pounded?
>
> With regard to MPI, what sort of interconnects will your execute nodes
> have? MPI wants lots of bandwidth between nodes and regular gigabit
> might not cut it (depending on your users' applications).
>
> --
>
> o--------------------------{ Will Maier }--------------------------o
> | web:.......http://www.lfod.us/ | email.........willma...@ml1.net |
> *---------------------[ BSD: Live Free or Die ]--------------------*
