opteron VS Itanium 2

Dr Eric Edward Moore eemoore at fyndo.com
Fri Oct 31 14:52:55 EST 2003


Mark Hahn <hahn at physics.mcmaster.ca> writes:

> there's a characteristic "shape" to spec results - which scores are 
> high and low relative to the other scores for a single machine.  not only
> does this include outliers (drastic cache or compiler effects), but
> points at strengths/weaknesses of particular architectures.  how to do this,
> perhaps some kind of factor analysis?

Well, being bored, I tried factor analysis on the average results for
the submitted specfp benchmarks at http://www.specbench.org/

The 5 factors with the largest eigenvaslues are:

Eigenvalue:    0.314116    0.353034    0.799331    1.432038   10.614996 
               2.22%       2.25%       5.70%      10.22%      75.82%

168.wupwise   -0.4134913   0.0241240  -0.1437086  -0.2757206   0.2715672
171.swim       0.0245451   0.0965325   0.3495143   0.1209393   0.2783842
172.mgrid      0.1122617   0.1365769   0.3273285   0.1332301   0.2839204
173.applu      0.0299056   0.0439954   0.4163242   0.1913496   0.2725619
177.mesa       0.4791260   0.4190313  -0.0949648  -0.3785996   0.2448368
178.galgel    -0.0489231  -0.5404192  -0.2464610   0.2391370   0.2648068
179.art        0.0646181   0.5095081  -0.4736362   0.6508958   0.1054875
183.equake    -0.5560255   0.0841426   0.0214064   0.1615493   0.2794066
187.facerec   -0.0402649   0.0446221  -0.2628912  -0.0557252   0.2897607
188.ammp       0.3993861  -0.3404615  -0.1456043   0.0359475   0.2832809
189.lucas     -0.2380202   0.0908976   0.0801927  -0.2140971   0.2842518
191.fma3d     -0.0326577   0.1661895  -0.1149762  -0.3148501   0.2774768
200.sixtrack   0.1950678  -0.1574121   0.2852895   0.2008475   0.2741305
301.apsi       0.1128198  -0.2379642  -0.3013536  -0.1224494   0.2782804

Pretty much all the specfp tests correlate with each other pretty
well, except for 179.art, which correlates...  poorly with the
others (it's correlation with 177.mesa is just 0.03).  So most of the
variation in the results is some sort of "raw speed" number, which has
near-equal weightings of all the tests besides 179.art.  

Next most important is whatever makes art so different from all the
others (maybe it's a persistent cache-misser, or maybe it's just the
easiest for vendors to tweak).

Not entirely sure what to make of the others.  There does seem to be
some commonality between 171.swim 172.mgrid 173.applu and 200.sixtrack
in the third biggest factor (plus a lot of whatever art isn't) that
could be important.  

The next two seem to mostly have something to do with whatever makes
177.mesa special.  

This is presumably all useless, but someone might be entertained :)

> regards, mark hahn.

-- 
Eric E. Moore
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list