> Scott Lampert wrote:
> > > Robert HYATT wrote:
> > > > I ran the matrix multiply benchmark... with 1, 2, 3 and 4
> > > > processors.
> > > > The numbers I got were 52 seconds, 59 seconds, 66 seconds and 73 seconds
> > >
> > > Hmmm... with four cpus the benchmark took 11 sec more then with 1 cpu ??
> > > One might suppose you meant the times to be :
> > > 52/1 1 cpu
> > > 59/2 2 cpu
> > > 66/3 3 cpu
> > > 73/4 4 cpu
> > > Cheers...
> > Umm.. I believe the difference is 23 seconds and I would imagine it to be:
> > 52 4 cpu
> > 59 3 cpu
> > 66 2 cpu
> > 73 1 cpu
> > But I could be wrong. ;)
> Yes, we were wrong, the difference is 21 seconds :)
> The way You look at it is even worse then the way I look at it :)
> What is the number of cpu's when you begin to lose ??
> Cheers..
>
depends on your definition of "lose". In the above benchmark, two
cpus interfere measurably, as the drop from 52 to 59 seconds total
time shows. Each additional processor interferes with the preceeding
number by adding an additional 7 seconds to *everybody*.
I didn't consider this horrible, because running 4 at once is stomping
through 160 megs of array in worst-case order, frying the memory bus
about as badly as it can be fried.
I have a computer chess program (Crafty) which I wrote, and which uses
multiple threads, and I don't see *any* degradation as new processes
are started, because it is a lot more cache friendly, and doesn't fry
the bus as badly. Real-world is somewhere in between, but for non-cray
sized arrays, it does well. If you have truly huge arrays, you take a
performance hit, but notice that 4 processors is still way faster
than 1, no matter what... 1 takes 52... 4 should take 62 each, but
see the total time stretch out to 73 total, which means I am doing
208 seconds of computing in 73 seconds, roughly 3X faster...
not bad when you factor in $13K for the total machine... which is
overkill for a workstation, with the fancy power supply monitor, fan
monitor, cpu temp monitor, ambient air temp monitor, case door monitor,
out of toilet paper monitor, etc.. :)
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu