Opteron?

[Martin Albrecht]

Hi,

I am planning to cut a new release of M4RI soon-ish part of which is the
rewrite of some of our base cases. As reported before I *think* it is
consistently faster than what we had before. I could confirm this on Intel
machines (see below) but I am having trouble testing this on Opterons. The
*.math.washington.edu network does have an Opteron (redhawk.*) but timings on
that machine are very very unreliable (they jump around a lot)

[I think partly because Leif (CCed) is running some large scale search on all
cores. He's nice about it, but for our purposes I am under the impression that
this is enough to screw up our timings.]

Hence, anyone here with an Opteron who wants to test the new M4RI release
against the old and/or wants to grant me a login so I can do it?

Cheers,
Martin

PS: Timings on barbecue (in Nancy):

Elimination:

m: 4096, n: 4096, m4ri-20120613: 0.067866 s, m4ri-next: 0.05677 s
m: 8192, n: 8192, m4ri-20120613: 0.423065 s, m4ri-next: 0.38568 s
m: 10000, n: 10000, m4ri-20120613: 0.758640 s, m4ri-next: 0.71572 s
m: 16384, n: 16384, m4ri-20120613: 2.914597 s, m4ri-next: 2.62982 s
m: 20000, n: 20000, m4ri-20120613: 5.106476 s, m4ri-next: 4.80096 s
m: 32000, n: 32000, m4ri-20120613: 19.508634 s, m4ri-next: 18.37055 s
m: 32768, n: 32768, m4ri-20120613: 19.661269 s, m4ri-next: 18.42793 s
m: 65536, n: 65536, m4ri-20120613: 137.418654 s, m4ri-next: 132.82949 s

Multiplication:

m: 4096, n: 4096, m4ri-20120613: 0.114194 s, m4ri-next: 0.09790 s
m: 8192, n: 8192, m4ri-20120613: 0.855474 s, m4ri-next: 0.77154 s
m: 10000, n: 10000, m4ri-20120613: 1.549899 s, m4ri-next: 1.41395 s
m: 16384, n: 16384, m4ri-20120613: 6.005692 s, m4ri-next: 5.40141 s
m: 20000, n: 20000, m4ri-20120613: 10.621466 s, m4ri-next: 10.40144 s
m: 32000, n: 32000, m4ri-20120613: 46.351486 s, m4ri-next: 42.85214 s
m: 32768, n: 32768, m4ri-20120613: 42.189423 s, m4ri-next: 37.32975 s
m: 65536, n: 65536, m4ri-20120613: 298.043383 s, m4ri-next: 265.78939 s

--
name: Martin Albrecht
_pgp: http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x8EF0DC99
_otr: 47F43D1A 5D68C36F 468BAEBA 640E8856 D7951CCF
_www: http://martinralbrecht.wordpress.com/
_jab: martinr...@jabber.ccc.de

[Martin Albrecht]

Hey, in the mean time Clément gave me an account on nice Opteron machines with
no load at the moment. So I am happy :)

On Thursday 08 Nov 2012, leif wrote:
> Martin Albrecht wrote:
> > Hi,(or both)

> >
> > I am planning to cut a new release of M4RI soon-ish part of which is the
> > rewrite of some of our base cases. As reported before I *think* it is
> > consistently faster than what we had before. I could confirm this on
> > Intel machines (see below) but I am having trouble testing this on
> > Opterons. The *.math.washington.edu network does have an Opteron
> > (redhawk.*) but timings on that machine are very very unreliable (they
> > jump around a lot)
> >
> > [I think partly because Leif (CCed) is running some large scale search on
> > all cores. He's nice about it, but for our purposes I am under the
> > impression that this is enough to screw up our timings.]
>

> Well, in my experience the scheduling on redhawk is extremely poor; not
> sure whether it's the kernel or the machine's [memory] architecture
> which is to blame. (I observed CPU times increasing by a factor of 6
> and more just because of a single additional background job of someone
> else which IIRC didn't even take up much memory, but for example rising
> the average number of running processes from 12 to 13 [while there are
> 24 cores on rh].
>
> I could of course suspend my (twelve) processes [which don't use much
> memory] on redhawk for a while; just let me know...
>
> FWIW, combinat[.math] is also an Opteron machine, although a newer one
> (four Opteron 6276, i.e., four 16-core Interlagos with 2 MB L2 per core,
> running at [just] 2.3 GHz).
>
>
> Cheers,
>
> -leif

Cheers,
Martin