[Beowulf] cluster softwares supporting parallel CFD computing

Vincent Diepeveen diep at xs4all.nl
Fri Sep 8 21:10:33 PDT 2006


You're assuming that you run 1 thread at a 2 to 4 core node or so?

This is not the reality.

See supercomputer report Europe. On average of machines in production more 
than 50% of the cpu's is on average in use and that
goes up to 70% on average at every given hour 24 hours a day, 365 days a 
year.

So even if you start your program as a single thread a node (note why make 
your program MPI when you have idle cores in this node),
even in that unlikely case, the odds is 70% some other job runs at that same 
node as well eating the other core(s).

Vincent

----- Original Message ----- 
From: "Eric W. Biederman" <ebiederm at xmission.com>
To: "Vincent Diepeveen" <diep at xs4all.nl>
Cc: "Ashley Pittman" <ashley at quadrics.com>; "'Bogdan Costescu'" 
<Bogdan.Costescu at iwr.uni-heidelberg.de>; "'Beowulf List'" 
<beowulf at beowulf.org>; "Daniel Kidger" <daniel.kidger at clearspeed.com>; 
"'Mark Hahn'" <hahn at physics.mcmaster.ca>
Sent: Saturday, September 09, 2006 3:49 AM
Subject: Re: [Beowulf] cluster softwares supporting parallel CFD computing


> "Vincent Diepeveen" <diep at xs4all.nl> writes:
>
>> How about the latency to wake up that thread again. runqueue latency in 
>> linux is
>> 10+ ms?
>
> That assumes you have a 100Mhz clock (default is currently 250Mhz) and
> you have something else running.  If you have something else running
> yielding is even more of a win because you get something productive
> done, when you would otherwise be idle.  Plus you the scheduler
> will give your process a higher priority when it wakes up because
> it slept.
>
> In practice I don't expect anything interesting will happen in that time
> interval.  So you should awake almost instantly.
>
> Eric
> 




More information about the Beowulf mailing list