We are interested in hearing your feedback on likes/dislikes of the clustering software management package you are using (Such as Scyld, OSCAR, NPACI Rocks, Score, etc.). Do you think that ease of installation is key, or is it worth a little bit of hassle to get installed in order to get a more robust platform?
Ease and completeness of installation are the keys as far as I'm concerned. Also, maintenance and ease of upgrade is a leading concern. If I can't install some library or other piece of support software across the nodes of the cluster with little effort, I might as well not have the management software package.
I've installed some Rocks clusters and have been impressed with the installation procedure. On OSCAR, I really appreciated the ability to run the same command on a set of cluster nodes with a single command. This allowed the setup and tear down of PVFS to be done quite easily.
For our purposes, we don't really have clusters that last for more than a few weeks, so I've not had the chance to see how well any one management software performs for an extended period of time. Thus, I really only have installation experience with these packages.
Rocks (and derived works) also have the ability to run commands across the nodes from a single command line. For those new to the list, here is the standard syntax for both packages:
cluster-fork "uname -a"
Rocks also has commands such as cluster-ps and cluster-kill.
cexec "uname -a"
OSCAR has one neat command called cpush that allows a file copy to each node. Of course, this command can be achieved with a detailed cexec or cluster-fork command too, but it ease of use is important.
On the subject of ease of installation, upgrade and maintenance, have you tried Scyld at all? How about some of the other ones, such as ThinOSCAR, HA-OSCAR, SSS-OSCAR, Clubmask, Warewulf, Clustermatic, SCore, OpenSCE, Mandrake CLIC, OpenMosix, and LSF?