Also, since most MPI type programs assume that the program is resident in the same location on all the machines, it just makes sense to distribute the home directories and have each machine have the same filesystem. It keeps things simple.
If you don't like the extra NFS trafic (it shouldn't be that expensive anyway unless you have a lot of users on a lot of nodes moving a lot of files around), I would suggest restricting access to just the head node, or to ssh from a non-cluster terminal to the head, rather than ditching the NFS shares. The whole reason the nodes don't have a lot of user functionality (xwindows etc) is that they really aren't intended to do anything besides work on the problems given to them by the head node. The head node doesn't have anything else to do besides act as a gateway, so give it something to do instead of the hard working nodes. Then NFS is just a convenience for your programs, and not a drain on resources. Of course, I have like 5 users tops, so this is very easy for me especially with a couple extra terminals. On Mon, 20 Sep 2004 15:26:20 -0700, Bernard Li <[EMAIL PROTECTED]> wrote: > Hi James: > > We actually have a NFS filer which serves /home and /scratch to the > cluster and it has worked out pretty good. > > I think generally if you are working in a 'grid' environment where you > are trying to use distinct resources from multiple clusters, then users > will be required to copy files back/forth. However, in a 'local' > environment, it would be really nice to be able to avoid that. > > I guess it depends on how much file I/O you are expecting, and how many > nodes you have - because obviously if all of your nodes started > generating a lot of NFS traffic, your NFS server may not be able to > handle it :-) > > Cheers, > > Bernard > > > > > -----Original Message----- > > From: [EMAIL PROTECTED] > > [mailto:[EMAIL PROTECTED] On Behalf Of > > James Wulf > > Sent: Monday, September 20, 2004 14:21 > > To: [EMAIL PROTECTED] > > Subject: [Oscar-users] Question about home and scratch directories > > > > > > All: > > > > I was wondering if people would mind sharing their > > implementation on home directories with regard to Oscar > > clusters with me. Most of my users have become accustomed to > > sitting down at any of our workstations and being able to > > login automagically and see the same home directory courtesy > > of NFS and NIS. > > > > But the default Oscar setup is pretty much self contained > > with home system-unique directories sitting physically on the > > head node and then NFS mounted to the compute nodes. I like > > that idea since it compartmentalizes the machine and keeps > > the network traffic between the nodes strictly at the > > business of cluster computing, as opposed to trying to NFS > > mount external home directories onto each node. > > > > In general, are most clusters setup so that users are > > expected to copy over their data, run it on the cluster, and > > then copy their results back? I was toying with the idea of > > mounting the external home directories onto the head node in > > a different location as a convenience to the users in copying > > the data over > > > > I was also wondering if it is common to setup local scratch > > directories on each node so that applications can write > > directly to disk as opposed to writing into the NFS mounted > > home directory. I'm not sure that everything we will be > > running is well behaved enough to just use tmp space. > > > > > > Thanks. > > > > James Wulf > > [EMAIL PROTECTED] > > > > > ------------------------------------------------------- > This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170 > Project Admins to receive an Apple iPod Mini FREE for your judgement on > who ports your project to Linux PPC the best. Sponsored by IBM. > Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php > _______________________________________________ > Oscar-users mailing list > [EMAIL PROTECTED] > https://lists.sourceforge.net/lists/listinfo/oscar-users > ------------------------------------------------------- This SF.Net email is sponsored by: YOU BE THE JUDGE. Be one of 170 Project Admins to receive an Apple iPod Mini FREE for your judgement on who ports your project to Linux PPC the best. Sponsored by IBM. Deadline: Sept. 24. Go here: http://sf.net/ppc_contest.php _______________________________________________ Oscar-users mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/oscar-users
