Re: Benchmark performance Under AIX
- To: mathgroup at smc.vnet.net
- Subject: [mg96638] Re: Benchmark performance Under AIX
- From: Bob F <deepyogurt at gmail.com>
- Date: Wed, 18 Feb 2009 04:22:22 -0500 (EST)
- References: <gn5ub5$gr8$1@smc.vnet.net>
The problem, I think, has to do with the math library routines of AIX
for the trig functions. You might look into getting newer math
libraries on your AIX system, but not sure if Mathematica would use
them as it depends on whether it is dynamically or statically linked.
Does anyone know?
The same problem exists on the Mac OS X version of Mathematica. On
version 6.0.3 of Mathematica the test #5 results for a 3 GHz Intel CPU
on Mac OS X were 10.3 seconds and for the same exact CPU on Windows
the result was 4.03 seconds, which is about 2.5 times better for that
one test on Windows. The overall result for Mac was 37.4 seconds (2.3
score) vs 30.4 seconds (2.84 score) for Windows. At the time I noticed
this, I looked at each of the functions in the #5 test and I recall
that the trig stuff was using most of the time. And I believe the
difference was due to the different older math libraries Wolfram used
on Mac OS X due to the newest Intel libraries not running on older
versions of Mac OS X. Looking at the performance of an AIX system on
the V6 benchmark it had a 60.13 sec for #5 and a 163.75 seconds
overall (on an AIX 5.3 Power 4 1GHz cpu), so the performance on this
one test was about 37% of all the time took, so bad results on this
one test really skews the overall results. You might talk to AIX
system provider and see if they know of ways to speed up math routines
on it.
The Mathematica v7 results for an Intel 3.0 GHz 5160 cpu were - for #5
test Mac OS X took 7.64 seconds and Windows took 4.03 seconds and for
overall Mac OS X 34.84 seconds (2.47 score) and Windows 30.39 seconds
(2.84 score), so the Mac OS X score has improved from V6 to V7, but I
have not examined the benchmarks in detail to know if the two are
exactly the same between the two versions of Mathematica, but I did
look at the test #5 results for the V6 and V7 benchmarks are they are
very very similar (V7 uses RandomArray[Real,...] and Exp[m1] and V7
uses RandomReal[] and e^m1) for 2 million exponentiations, sins and
arctans. Since V7 of Mathematica requires Mac OS X 10.4 or newer, I am
guessing they were able to use newer Intel math libraries (and hence
faster than previous versions).
These newer Intel math libraries gave about 30% improvement in just
these trig functions (but only about 7% overall) on Mac OS X. I
rather doubt that Wolfram did much improvement in the way they do Sin
and ArcTan and Exp but it is possible. I think V6 of Mathematica
required Mac OS X 10.2 or newer (not sure).
-Bob
Coleman, Mark wrote:
> Greetings,
>
> I've been doing some benchmark testing of a new IBM Server running AIX
> that we use for Mathematica. In particular I am using the
> MathematicaMark6 tests that come with Mathematica v6.03. I've gotten
> a puzzling result and I was wondering if anyone else on MathGroup has
> seem a similar issue.
>
> Briefly, we have Grid-Mathematica installed on an 6-CPU p500 series
> IBM server (the machine uses PowerPC Power6 CPUs at 4.2 Ghz each) with
> 32Gb of Ram, running ver 5.3 of 64bit AIX (IBM's Unix). I've been
> comparing the performance of this machine versus my standard 2Ghz
> Laptop running Windows 2000. (Note: I am not making use of any grid
> capabilities on the server)
>
> Given the configurations of each machine, I would expect the server to
> be as much as 2x as fast as my laptop, given the difference in CPU
> speeds if nothing else. But this is not the case. The server results in
> a MathematicaMark6 score of 1.37 versus 1.67 for my laptop. I've traced
> the difference to Test #5 (Elementary Functions). The benchmark code for
> this test is:
>
> Test 5:Elementary Functions
>
> Module[{m1,m2},Timing[SeedRandom[1];m1=RandomArray[Real,{},{2000000}];
> m2=RandomArray[Real,{},{2000000}];
> Do[E^m1;Sin[m1];ArcTan[m1,m2],{30}]]]
>
>
> This code takes 6.06 seconds on my laptop and 26.8 seconds on the
> server!
>
> I'm wondering if there is a known issue on AIX (or UNIX systems more
> generally) that might account for this?
>
> Thanks,
>
> Mark