BEOWULF cluster hangs

[Donald Becker]

On Thu, 26 Sep 2002, Jeff Nguyen wrote:

> 2.4.7 kernel performs horribly on the dual Xeon platforms because it does not 
> support IRQ balancing and DMA transfer on the IDE bus. If you are using an
> IDE disk on the nodes, disk performance is very slow which might affect your
> application.

Very good points.  In this cases it's likely just the VM subsystem
problem because of the large memory and time element.

An IDE problem presents itself as apparently heavy disk use, especially
when short of memory.  Even before using swap, demand paging increases
disk usage.

But IDE problems show immediate bad performance, and here initial
performance is good. 

No IRQ balancing shows up as asymettrical CPU use, and usually
unpredictable short-term performance.  OTOH, a _bad_ implementation of
IRQ balancing causes cache thrashing, which results in overall bad
performance that can't readily be tied back to the cause.

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