newbie: 16-node 500Mbps design

Mark Hahn hahn at coffee.psychology.mcmaster.ca
Mon Aug 28 17:12:09 PDT 2000


> And many, many people would not find it dishonest to prefer 2.2 over 2.4,
> because although 2.4 has received much more testing than any release of 2.3,
> 2.4 is still new. No pretending required; it's a simple fact. When 2.4 hits
> version 2.4.5 or so and starts shipping in major distributions, then it will
> be considered mature.

the labels are problematic: 2.2 is called "stable", but is not more 
stable (robust).  multiple revisions make a release "mature",
but does not imply improvement of any kind; nor does more testing.

the only truely relevant benchmark for your cluster is *your* code.
similarly, other people's random testing (of 2.2 or 2.4) is only
weak evidence of how it'll run on your cluster.  if your application
does little IO and makes few syscalls, and is running happily under 2.2,
then don't bother looking at 2.4.  reality is that 2.2 is merely the state of
Linux circa late summer 1998.  by definition, it lacks the 2 years of
improvements that 2.3/2.4 benefits from.  many of those improvements 
are directly relevant to  clusters, especially the major changes 
to SMP locking, VM, networking.  many of the changes during 2.3 were
redesigns to solve problems inherent to 2.2, not added functionality.

regards, mark hahn.








More information about the Beowulf mailing list