Cluster Upgrade Strategy

[Donald Becker]

On Thu, 17 Oct 2002, Ed Hall wrote:
> I have some fairly basic questions about upgrading/expanding a existing
> (less than 1 year old) Linux cluster with additional nodes. The cluster
> is used in a University environment for both serial and parallel processing.
> 
> Is it better to buy older cpus that are the same as on the existing nodes
> to keep the cluster homogenous so that its easier to administer and use?
> Or does it make more sense to buy newer (and possibly fewer) cpus with
> greater compute capacity at the expense of having a uniform set of nodes
> in the cluster? If the only alternative is to buy newer nodes, should they
> be used as a separate cluster rather than combining them with an existing
> cluster of older nodes?

A good cluster system should make this a question purely about
scheduling and your application's internal load balancing, not about the
processor speed ratio or memory in each system.

If your applications operate in lock-step and are sized to exactly the
cluster size, different speed nodes will result in decreased efficiency
and you should set them up as two different logical cluster, or perhaps
just two different scheduling domains.

If your applications have dynamically allocated work, just put all of
the nodes in a single cluster.  They will be much easier to manage.

In the Scyld Beowulf system (well, specifically the professional
editions) we ship six different kernel variously optimized for
Pentium-4, Athlon, and uniprocessor/SMP.  And with the current system
the whole cluster runs a single common kernel.  In this case there
is a motivation for keeping the processor family the same.

-- 
Donald Becker				becker at scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Scyld Beowulf cluster system
Annapolis MD 21403			410-990-9993