[mpich-discuss] Disk IO on multi-core MPI?

[Dorian Krause]

Thomas Blom wrote:
> Hi,
>
> This post may belong somewhere else, but I'm having a hard time 
> locating good information on disk i/o as it relates to MPI software 
> running on multi-core machines.
>
> I'm in a position to influence what type of node configuration we use 
> for a new cluster doing bioinformatic-type work == lots of disk I/O.  
> It occurs to me that multiple MPI processes running on the same 
> multicore node may not prove very effective if they are constantly 
> invalidating each others' disk cache etc... 
>
> I'm wondering about the idea of instead of a single disk for storage, 
> using multiple disks (e.g. 1 per core) to avoid these kinds of 
> collisions.  Is this a common practice?  Certainly it complicates the 
> underlying code a little, since core N now needs to read/write a 
> different disk than core M on the same machine.
>
> There must be well-written papers about this very issue out there, but 
> I haven't found any yet.
>
> thanks,
> thomas blom, austin tx
Hi,

If you talk about a whole cluster of multicore processors you will have 
to use a shared filesystem anyway?! In this case I wouldn't matter 
because disks aren't bound to nodes anymore. Depending on your 
applications I would definitely prefer a parallel filesystem (PVFS, 
Lustre, GPFS, ...) instead of e.g. NFS. It can make a huge difference!

Regards,
Dorian