[lug] Cluster Mfr Recommends

[Sean Reifschneider]

On Mon, Mar 14, 2005 at 09:25:12PM -0600, Michael J. Hammel wrote:
>I worked at RLXTechnologies the past year.  They made blade systems. 
>But they axed their hardware biz and now just do blade management

I've been looking at blade systems for the last several years, and the
biggest problem I've had with them is their price.  While it's not hard for
a 1U system to get up in the $2k range, there's a fairly good selection
down under a grand.  Blade systems, on the other hand, start around $3k per
system.

There are some benefits, like higher density and better or worse
management, etc...  There are some less expensive options if you're ok
going with <1GHz CPUs.

>While there, I got to work with SuperMicro 1U rackmounted systems. 

We've used a lot of SuperMicro 1U systems in the clusters we've set up for
our clients.  They're fairly solid machines, though I wish more of the
low-end ones supported ECC RAM.

>Another Linux vendor is Pogo Linux up in Seattle.  Haven't worked with

The Pogo boxes I've seen seem to just be re-branded MSI systems, so I'm not
entirely sure why I'd go with Pogo over buying MSI direct.  They may add
something on the software testing side, I'm not sure.  I'm not saying I
wouldn't go with them, just that I don't know what the benefit is.

>If a single cluster is "6 to 12 computers" then a blade chassis sounds
>like a good fit for you.  Blades are nothing more than lots of

If you are interested in the price premium, or the management functionality
available from them is something you need.  It's unlikely that at 6 to 12
machines in a rack you really NEED the density benefits.

For HA applications, the blade setup probably isn't what you want though.
You don't want to put all your eggs in one chassis, and it can make the
Linux-HA "STONITH" hard to do as (AFAIK) all of the existing code is
designed to work with power strips or UPSs to kill the other node (and
ensure no corruption), not blade chassis.

We have built a new plugin called "external" for STONITH which makes it
much easier to develop external STONITH plugins that can do things like
power off a system in a chassis using it's appropriate control mechanism,
which is probably what you want to use there.

>reduces to that kind of problem).  High availability software will
>probably come from a 3rd party vendor or maybe the Linux-HA project. 

tummy.com has a fair bit of experience doing Linux-HA cluster setups.  :-)
IBM is employing Alan Robertson, the Linux-HA lead, so I'm sure they have
resources as well.

>RLX had some cluster add ons but I think it was just a wrapper around
>some beowulf stuff.  I never saw it in action so can't really say what
>it did.

That sounds mostly like compute cluster as opposed to HA cluster related
though.

Sean
-- 
 "No early worm is giving ME the bird!" -- Bullwinkle J. Moose
Sean Reifschneider, Member of Technical Staff <jafo at tummy.com>
tummy.com, ltd. - Linux Consulting since 1995.  Qmail, Python, SysAdmin
      Back off man. I'm a scientist.   http://HackingSociety.org/