hsv@lanl.gov (Henry S Vaccaro) (04/14/90)
In article <1990Apr5.215435.8158@ico.isc.com>, rcd@ico.isc.com (Dick Dunn) writes: > jfh@rpp386.cactus.org (John F. Haugh II) writes: > > > Apparently Dhrystone 1.1 isn't such a bad benchmark since a recent > > review of IBM's System/6000 included SPECmark numbers which bear > > out IBM's performance claims. > > Oh, sure they do...if you can believe that 27.5 is pretty close to 15.8! > I'd say the SPEC figures absolutely deflate any Dhrystone 1.1 claims. > > The SPECmark for the 20-MHz 6000 is 22.3 x the "ref machine" (VAX 11/780), > or"22.3 MIPS" in layman's terms. So even there, a 27.5 MIPS claim is high > by more than 20%...but that's not the real point. > > *Remember* that Dhrystone is an integer benchmark, and SPEC is 6/10 > floating point. If, instead, you use the "integer SPEC number", which is > the geometric mean of the 4 integer SPEC benchmark ratios, the number for > the 20 MHz 6000 is 15.8 MIPS. The 27.5 MIPS claim is almost a 75% over- > statement of the machine's integer performance. > > Make no mistake, the 6000 is still a fast machine, especially so if you're > using it for a lot of floating point. But it's just NOT 27.5 integer MIPS > by any stretch of the imagination. Well, just which are the right Dhrystone numbers anyway. Hereere a few choices for the RS/6000-530: $ cc -O -U__STR__ -o dhry1-1 dhry1-1.c $ dhry1-1 Dhrystone(1.1) time for 500000 passes = 8 This machine benchmarks at 61804 dhrystones/second $ $ cc -O -o dhry1-1 dhry1-1.c $ dhry1-1 Dhrystone(1.1) time for 500000 passes = 7 This machine benchmarks at 64432 dhrystones/second $ $ cc -O -Q -o dhry1-1 dhry1-1.c $ dhry1-1 Dhrystone(1.1) time for 500000 passes = 6 This machine benchmarks at 82918 dhrystones/second $ (Note: HZ is set to 100 in dhry1-1.c) I've just finished my own tests on a mostly integer code I use here. The mackine above is 9-10 times faster than an IBM RT Model 125. That machine is roughly "3 MIPS", and runs the code 5-7 times faster than our old 780 could. Does this mean that the RS/6000 is 5x9 to 7x10 VUPS? Another datapoint: our VAXSTATION 2000's yield: $ cc -O -o dhry1-1 dhry1-1.c $ dhry1-1 Dhrystone(1.1) time for 50000 passes = 38 This machine benchmarks at 1311 dhrystones/second $ Hank Vaccaro hsv@lanl.gov rt1!hsv
mash@mips.COM (John Mashey) (04/15/90)
In article <48588@lanl.gov> hsv@lanl.gov (Henry S Vaccaro) writes: ... >I've just finished my own tests on a mostly integer code I use here. >The mackine above is 9-10 times faster than an IBM RT Model 125. That >machine is roughly "3 MIPS", and runs the code 5-7 times faster than our >old 780 could. Does this mean that the RS/6000 is 5x9 to 7x10 VUPS? Could you describe these benchmarks? And also, give the version of the VAX11/780? Including the compilers? It is very easy to get different numbers depending on the versions of these things. If some of the benchmarks are real applications, or close, could any of them be posted so people could try them out on others? The reason I ask, is that the SPEC integer codes: a) Are all real programs. b) Are compared against a current 11/780 with current compilers. and the 530 gives relative performance ratings of: 17.8, 20.7, 19.8, 23, for geometric mean of 20.2. whereas your calculations would indicate 45-70. Hence it would help to publish enough more data for us to figure out why there is a factor of 2-3.5X difference. Note, as alwaqys, that Dhrystone is one of the best possible programs to use to prove how much better than a VAX you are, for reasons covered many times in this group. -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: {ames,decwrl,prls,pyramid}!mips!mash OR mash@mips.com DDD: 408-991-0253 or 408-720-1700, x253, or 408-524-7015 USPS: MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086