On Tue, 2007-10-23 at 15:48 -0400, Robert G. Brown wrote: > I expected the latter, actually. Scyld may be a similar sort of model > as well. I was more curious about how many people on list are running > clusters that are currently configured to run nodes set up for a minimal > kernel or specialized kernel with little multitasking. > > This fits in the tradeoff I described. Very large clusters shift the > marginal benefit curve for the work and expertise required to run a > cluster where the nodes are basically very stupid and very specialized. > Easiest is to just make a kickstart file or package list for a "rich" > node that has more than the bare minimum set of stuff installed, and > that can still (for example) run logins and ssh's in userspace (if only > for root and administrative users). > > rgb
I would have to agree with this observation, as we are currently migrating from a Scyld environment to a standardized Red Hat (insert you favorite distro) environment. The issue is not that Sclyd has performed poorly, rather that we find it inflexible when we need to do a variety of different tasks with the same cluster resources. If we only ran one or two MPI models we would undoubtedly find the minimalistic approach a benefit in both speed and management. As it it, we run almost no parallel or MPI codes. It seems more and more often we are temporarily re-tasking cluster nodes to run Windows to run models (Windows modelers found out we have fast boxes). In the end, the ability to login to the nodes, use standard tools like kickstart and pxeboot to provision the OS , be it Windows or Linux, and integrate the various systems with different installs into the same Infiniband and Gigabit network using the same storage has proven far more valuable than the benefits from running a minimal kernel. ~Glen > > > > _______________________________________________ > > Beowulf mailing list, [email protected] > > To change your subscription (digest mode or unsubscribe) visit > > http://www.beowulf.org/mailman/listinfo/beowulf > _______________________________________________ Beowulf mailing list, [email protected] To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
