[Beowulf] Windows cluster

[Robert G. Brown]

Nicholas M Glykos writes:

> <snip>
>> You should contact the Cornell Theory Center folks.  They have (one of 
>> the very few) windows clusters around, and lots of experiences 
>> running/building/managing them.
> </snip>
> 
> I am not qualified to pass judgement on CTC's choices, but by the look of 
> it (see http://www.tc.cornell.edu/CTC-Main/Services/Support/ 
> general_topics.cluster_status.info.htm) they are busy in ways rather 
> uncommon for HPC :
> 
> ...
> DOWNTIME: Wednesday September 14, 8am-5pm
> REASON: Apply Microsoft Security Patches.
> ...
> Posted Friday, 19 August 16:25
> The computer worm that was released into the Internet around August 13th 
> affected the stability of the CTC Windows domain. ...
> ...
> DOWNTIME: Wednesday August 17, 8am-5pm
> REASON: Install new image on V2 nodes. (Windows 2003, sp1). See details 
> below. Apply Microsoft Security Patches.  
> ...
> DOWNTIME: Wednesday June 29, 8am-5pm
> REASON: Apply Microsoft Security Patches.

What a hoot!  I'm going to save this little snippet, as it should just
plain go into the list faqs for automated response to the question "How
to I build a Windows cluster?".

In the meantime, most linux distros these days update all systems
nightly, applying all security "patches" and other updates smoothly and
completely transparently to the user via e.g. yum.  Cluster nodes
routinely run for stretches of -- as long as you want them to, limited
by the rate of HARDWARE failure.

And you KNOW that the Cornell cluster, funded by people with a deep and
abiding wish to prove the viability as a clustering platform, uses
gold-plated hardware...

I'm actually surprised and a little bit appalled that even in the last
half of 2005, Windows is STILL so unstable that it requires once a month
downtime for updates.  Of course, any time you update ANYTHING on a
linux box you have to reboot -- on a linux box you only have to reboot
(typically) if you update into a new security-patched kernel, and those
are rare indeed (and usually unnecessary to apply/reboot in any hurry in
a firewalled cluster).

   rgb
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20050909/b2d446ae/attachment.bin