palmer@ncifcrf.gov (Thomas Palmer) (03/31/89)
We received a DEC 3100 demo unit today and I've run some benchmarks comparing it to our 4D/70GT. The results are somewhat perplexing. On a floating point test (just raw +,-,*,divide) the 3100 performs at about 2.2 mflops; the 4D/70GT at only 1.02 mflops. With a Dhrystone the 3100 produced 24590 dhrystones/sec; the 4D/70GT only 10204. Our Cray only churns out 14705 dhrystones/sec! What gives here? They have the same chip set (R2000 plus R2010 fpa, right?) The compilers are the same. Are the clock speeds the same? Is DEC using more cache? Any ideas? -Tom Thomas C. Palmer National Cancer Institute Supercomputer Facility c/o PRI, Inc. Phone: (301) 698-5797 PO Box B, Bldg. 430 Uucp: ...!uunet!ncifcrf.gov!palmer Frederick, MD 21701 Arpanet: palmer@ncifcrf.gov
jim@kaos.Stanford.EDU (Jim Helman) (04/01/89)
> With a Dhrystone > the 3100 produced 24590 dhrystones/sec; the 4D/70GT only 10204. The copy of dhrystone.c that I have defines the clock tick rate, HZ, to be 60. On the 4D machines this should be 100 (see /usr/include/sys/param.h). Using -O3 optimization, our 4D80 (16.7MHz MIPS) clocks in at 29250 dhrystones/sec. However, the whole question is somewhat moot, since the dhrystone is a pretty worthless benchmark. So now ya know. -jim
khb%chiba@Sun.COM (chiba) (04/01/89)
In article <8903311650.AA03992@aero4.larc.nasa.gov> blbates@AERO4.LARC.NASA.GOV (Bates TAD/HRNAB ms294 x2601) writes: > > I am not familiar with the Dhrystone test, but if it has a >lot of integer math, that will slow it down on the Cray. On the >Cray's real math is faster than integer math. Don't ask me why, >it is a known fact. The Cray's are optimized for real math and >they don't care about integers. >-- Dhrystone is _purely_ an integer test. It is basically worthless for predicting performance of real code, but since a wide variety of periodicals (Byte, et al) and netfolk insist on running it, and publishing the results manufacter's of small computers do pay attention (some more so than others, note the flap in comp.arch re: intel 860 performance :>). Cray customers benchmark the machines the old fashion way....they take their real codes to cray (or reasonable subsets thereof) and have them run them. (so, btw does Sun). This is the only way to really judge a computer, have those who know how to tickle the beastie run your real codes and faithfully report what happened, and how. Those who have small budgets rely on the press/net to give them the best info they can .... and thus Dhrystones lives on...and on.... Cray's actually have a good balance between integer and FP....for the class of applications for which the machine was designed (i.e. numerical math...so integers are for do-loop indicies and address computation only). If your application is different (say symbolic computation or certain types of circuit simulation) you may find that a very different machine is what the doctor ordered. Well enough pontificating for now. I return you to your regularly scheduled sgitalk :> Keith H. Bierman It's Not My Fault ---- I Voted for Bill & Opus
msc@ramoth.SGI.COM (Mark Callow) (04/01/89)
In article <771@fcs280s.ncifcrf.gov>, palmer@ncifcrf.gov (Thomas Palmer) writes: > > What gives here? They have the same chip set (R2000 plus R2010 fpa, > right?) The compilers are the same. Are the clock speeds the same? > Is DEC using more cache? Any ideas? The clock speeds are different. -- -Mark
blbates@AERO4.LARC.NASA.GOV (Bates TAD/HRNAB ms294 x2601) (04/01/89)
I agree, the only reliable way to bench mark machines is to take real programs and run them on different machines. I bet if someone changed the Dhrystone test so that it used ALL reals instead of integers, it would run faster on the CRAY, than the integer version on the CRAY. Could someone send me a copy of the Dhrystone test, if it isn't too big? Thanks -- Brent L. Bates NASA-Langley Research Center M.S. 294 Hampton, Virginia 23665-5225 (804) 864-2854 E-mail: blbates@aero4.larc.nasa.gov or blbates@aero2.larc.nasa.gov