Brian Barrett wrote:
>On Sep 17, 2005, at 11:22 AM, Dave Hagedorn wrote:
>
>
>
>>Just wondering if anyone has noticed any performance differences in
>>runnnig lam off of an NFS mount as opposed to running locally.
>>
>>We've got a 14-node cluster set up where I work. There is also a
>>storage server with a RAID array. All of the nodes and storage server
>>are connected with gigabit ethernet through a single switch. We've
>>stored lam + our program (LS-DYNA) on an NFS mount which all of the
>>nodes access. In addition, our home directories are mounted via NFS,
>>so all output fies from LS-DYNA gets sent across the netowrk. These
>>files typically amount to 2 GB or so, usually written over the course
>>of a days or more.
>>
>>Does anyone think this setup could cmpromise overall performance? We
>>haven't tried intstalling anything locally - the NFS setup is really
>>convenient.
>>
>>
>
>There's always going to be some performance hit when all your disk I/
>O and MPI communication go over the same network. I'd be more
>concerned with home directories (and therefore your LS_DYNA data)
>over NFS than I would be about some "site-local" packages that live
>in NFS on the cluster. Your startup times for LAM might be a bit
>lower using NFS instead of local disk, but on a 14 node cluster, it
>should be more than manageable. And once the programs are up and
>running, they should mostly be mapped into memory and not be a
>problem. If you are running really close to your physical memory,
>you might see extra NFS traffic as parts of LAM and your application
>get flushed out to disk, but that's about it.
>
>I'm not sure how big LS-DYNA's input and output data files are, but
>they are surely bigger than LAM's binaries. So I'd be inclined to
>worry more about them than anything else.
>
>
>
With the option -p=<pfile>, LS-DYNA MPP provides for designation of a
global and a local filesystem. Generally, nfs mount is satisfactory for
the global, if it is mounted only for the cluster. Local does in fact
need to be a local disk on each node. It's moving in the direction of
needing a performance RAID for an nfs global filesystem. With a cluster
of this size, attention to ethernet card driver parameters is important.
|
