Re: First Beowulf cluster. Some pointers?
stramiello (firstname.lastname@example.org) said:
> I think you're going to need a larger hdd than that. To have an NFS-root
> diskless cluster, you really need separate copies of the OS storied on the
> master node for each of the slave nodes.
Try to avoid NFS at all costs. (disclaimer, i'm pretty much just
retyping what i've seen yesterday on the beowulf mailing list) NFS
works well in a workstation environment where a user does some IO but
is generally idle a lot of the day. NFS in a beowulf is not so good
a choice because it's slow (especially without the 2.4 NFS backport)
and has cache consistency problems. Bootup will suck, since 15 nodes
are going to try to access the NFS server at once. (not that it can't
handle it, but serving 15 root partitions takes time: time you could
use doing computations)
a lot of people do have an NFS home directory across the cluster, but
the users (if they are looking to maximize their cycle usage) work
from a /scratch or some other local filespace.
Rob Latham: linux A-Team Bethlehem, PA USA
GPG Fingerprint: EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897