[UCI-Linux] Lab and Cluster Provisioning Discussion
harry.mangalam at uci.edu
Thu Jan 21 18:22:51 PST 2010
<Please forward to anyone who might be interested..>
(and sorry for any duplicates)
Some of you on the BDUC list may have noticed some Perceus vs ROCKS
postings from me over the past couple of days. This is a result of
the (perhaps impending) consolidation/collision of MPC and BDUC into
one larger resource, as well as a personal interest in Thin Client
(TC) computing and its various models. My interests are Linux-based
TCs, but many others assume it involves Windows.
It turns out there is considerable overlap in cluster computing and TC
computing and many of the technologies are applicable to the other:
- net/PXE booting
- provisioning of large number of clients
- selection of OS images (client or server-side)
- virtualization of various kinds
- updating/upgrading across clients
- balancing OS and apps in the image
- how thin the client should be (anorexic, thin, slim, fat, obese)
- where the code should run (on client, on single OS on server,
on virtualized OSs on server, on virtualized OSs in a cloud), etc.
I set up a Forum for members of the UCCSC to discuss TC computing, but
mostly it's been me ranting to myself:
If you want to read or sign up and contribute, I'd appreciate your
At any rate, for these reasons, it may be useful for those of us who
admin clusters and labs to have a discussion about the state of
cluster / cloud / TC computing and what it means for UC, both in the
scientific cluster space as well as in the lab/dept provisioning
Despite my tendencies towards Linux, I'm happy to hear about and
consider successes and issues with Windows and even Macs.
Some of the discussion topics (obviously Linux slanted) could be the
cobbler - system imager/provisioner, sort of like ghost;
puppet-aware; works with 'koan',
puppet - system config management (already being used by ICS, I think)
koan - image chooser; choose your image from a cobbler-supplied
lib of images.
func - like 'clustexec.pl' (parallel cluster commandline) but
better, stronger, more configurable, able to leap tall
buildings in a single bound
genome - server definition and replication tool riding on top of
cobbler, puppet, and func. Can use to build clouds of
perceus - stateless (OS does not hit disk) provisioning tool for
clusters, other groups of clients, such as departmental
nodes, labs, etc, using a hybrid provisioning model
(netboot of core OS, user-level shared apps, libs, etc
instalinux - <http://www.instalinux.com/> web page that will generate
a basic Linux OS image for you, based on your input.
rocks - cluster management tool from SDSC; probably largest number of
installations, tho perseus probably has the largest
deployments (10s of thousands of cores).
NB: All of the above tools were developed on and still seem to be
pretty RHEL-centric. Debian/Ubuntu support is coming along but
schedulers: SGE vs Torque/Maui vs others
Other - most distros are essentially cluster-ready now. Torque and
slurm are available from the Ubuntu repos. SGE is Open
Source; all you need is one of the schedulers and a way
to script some admin stuff and away you go.
I've been playing with Perceus and it is pretty neat, especially being
to easily provision a subset of nodes with a new config so you can
play with a new cluster config without committing the whole cluster.
On the other hand, I don't have a lot of experience with it, but I do
like that the guy who wrote it (Greg Kurtzer) is available in the
blink of an eye, it seems.
I'd like to hear how ROCKS is working out on GP and any other clusters
its being run on, as well as any thin client implementations on
Harry Mangalam - Research Computing, NACS, Rm 225 MSTB, UC Irvine
[ZOT 2225] / 92697 949 824-0084(o), 949 285-4487(c)
MSTB=Bldg 415 (G-5 on <http://today.uci.edu/pdf/UCI_09_map_campus.pdf>
Why would any society rely on a health care system that has a stated
fiduciary responsibility to extract as much money from its customers
while giving them as little care as possible?
More information about the UCI-Linux