From: <er...@he...> - 2004-08-10 16:15:20
|
On Sat, Jul 24, 2004 at 04:46:10PM -0500, Luke Palmer wrote: > Hi all, > > I find myself continually adding to my bproc libraries list, which is by > now huge. I was considering just making all my libraries available to > nodes via an NFS mount. Can anyone think of a reason why that would be > bad? The stuff we do here is very long running, so the increased > library loading time would be negligible. NFS isn't unreasonable if your cluster is small enough. I don't try to avoid it as much as I avoid depending on it. Process startup (especially MPI) creates a situation where potentially every node in the system will hit the NFS server at the same time. This can go badly in very large configurations. We've also seen coherency problems with network file systems (NFS and others) in very large systems. On the other hand, if you're not seeing any problems like that then I can't see any reason not to use it. - Erik |