[Beowulf] cluster softwares supporting parallel CFD computing

[Ashley Pittman]

On Wed, 2006-09-06 at 11:10 -0600, Eric W. Biederman wrote:
> "Daniel Kidger" <daniel.kidger at clearspeed.com> writes:
> 
> > Bogdan,
> >
> > Parallel applications with lots of MPI traffic should run fine on a cluster
> > with large jiffies - just as long as the interconnect you use doesn't need to
> > take any interrupts. (Interrupts add hugely to the latency figure)
> 
> I know a lot of NICs misplaced interrupt mitigation adds to the latency figure.
> I know people who are serious about optimizing their hardware for mpi traffic
> poll for interrupts.
> 
> There is fundamentally more work to do when you take an interrupt because
> you need to take a context switch.  But cost of a context switch is in
> the order of microseconds, so while measurable taking an interrupt should 
> not dramatically your latency numbers.
> 
> Do you have any measurements to show that interrupts add hugely to latency
> figures, as opposed to a poor implementation of interrupts.

I think Daniel was talking about supercomputer networks and not
ethernet, on the first QsNet2 machine I have to hand latency without
interrupts is 2.72uSec, using interrupts it is 7.20uSec.  One
fundamental difference between these two measurements is that when you
use interrupts the kernel has to get involved, without interrupts it
doesn't so you don't just have the interrupt but also a extra syscall.

I know of people who have tried to tune the Hz rate down as low as 10,
I'm fairly sure they wrote a paper about it.

Ashley,