[nycbug-talk] BSD Cluster Filesystem Roundup
akosela at andykosela.com
Wed Feb 25 02:32:13 EST 2009
Miles Nordin <carton at ivy.net> wrote:
> >>>>> "ak" == Andy Kosela <akosela at andykosela.com> writes:
> ak> The main goals would still remain the same: single root,
> ak> single process space, single I/O and IPC space, virtual
> ak> network address space
> oh, I suppose I could be wrong, but no, I don't think many 100 node
> clusters are doing that! They are batch not timesharing systems, and
> are more interested in how to run one userspace program that is
> specifically written for the cluster, and run it with low IPC latency
> even at the cost of ponderous interfaces (RDMA), and in having a batch
> job scheduler that can mark off commit-points and restart jobs
> dispatched to nodes that crash.
You are writing here about HPC, right? SSI clusters are completely
different and more interesting from my point of view.
> Not interested in how to make 100 cheap systems feel like a single old
> expensive Unix dinosaur like the college shell servers with 500
> undergrads logged into them all running 'elm' or 'slrn' or something
> timesharing jobs with erratic performance requirements that justify
> this idea of ``migration'', and because they are different users
> minimal > IPC aside from the shared filesystem and maybe pipes. I
> think this idea of clusters is more nostalgic than relevant:
IMHO those "community" boxes is what really made Unix fun back in the
days...and VMS too. An idea to have a single, very stable, reliable
system with many nodes is very exciting. In a way Internet as a whole
is such a collective single system now.
> sounds neat though, and I hadn't heard of OpenSSI before only of Mosix
> so I'm glad for the reference. I didn't know something like it was
> part of Tru64. I wonder if you get the Tru64 cluster package in the
> copy of Tru64 you get from the ``developers and enthusiasts'' program?
> probably not, but I have a copy of that, and I have two or three
> alphas (though they're busy running NetBSD right now).
OpenMosix is dead now, the project is called LinuxPMI nowadays. Not
sure about the copy of TruCluster in hobbyist license program, but Tru64
is really dying -- HP should port it to HP-UX which now has only HP
MC/ServiceGuard option and is really only an HA active/passive cluster.
More information about the talk