performance:
P6-200 AXP 21164-300
test1 150 MegaFlops 163 MegaFlops
test2 125 MegaFlops 0 MegaFlops (as in zero - endless loop ?)
test3a 90 MegaFlops 90 MegaFlops (short vectors)
test3b 50 MegaFlops 35 MegaFlops (long vectors)
test4 29.7 seconds 23.4 seconds
test5 8.4 seconds didn't run
test6 19.7 seconds didn't run (21064A-275, NT/DEC Fortran 33 s)
test1 is a loop calculating distances between 3-D vectors
test2 is a loop rotating/translating 3D-vectors
test3 is a loop of assorted vector operations
test3a is vector length = 500 - 10,000
test3b is vector length = 50,000 - 500,000
test4 is a mini molecular dynamics benchmark
test5/6 are crystallographic least squares refinements of large molecules
test5 and test6 didn't run, because one of the involved programs aborted
with an I/O error (end-of-file) - the EXACT same program with the EXACT
same data files does of course run on Intel linux (I tested it once again
after I came home with the same files that were used on the Alpha, on a
linux P5-90 and of course they ran without any problems, 38 sec. on P5-90
- for test5)
test4, 5 and 6 are Fortran (6 with some C [not in the offending program])
compiled with f2c+'gcc -O3' on the AXP, f2c+'gcc -O1' on the P6.
test1, 2 and 3 are in C - they are small, so anyone who would like to run
them on an Alpha, can have them right away (and maybe find out why test 2
hangs).
NOTE that all my tests emphasize floating point performance, in which the
AXP is supposed to shine.
The test machine has an on-board DEC ethernet, NCR53C810 SCSI and
Diamond Stealth video card, 2 GB IBM SCSI disk.
My impression is that f2c/gcc/libc on the Alpha is not what it should be.
Theses tests are very depressing, as it was nice to see a fully functional
Linux system running on it.
I DO HAVE the money to buy such a machine (and do so for quite some while)
with either 64 or 128 MB memory and 6 GB disk.
Could some AXP guru please comment on the above, before I either
1) jump out of the window
or
2) take the money and buy two P6-200 for it (which will be the same price).
Open and prompt response will be much appreciated, as I don't want the money
left unspent much longer. Any ideas what I might have been doing wrong -
any soul that could provide me with some serious floating point benchmark
results, run on botth a 21164-300 and a P6-200 under linux.
Arno
----------------------------------------------------
e-mail: paehler@atlas.rc.m-kagaku.co.jp
tel: +81-45-963-3868 (or 3156)
fax: +81-45-963-4206 (or 3974)
home: +81-45-985-1585 (also fax)
Dr. Arno Paehler, Senior Research Scientist
Research Center, Mitsubishi Chemical Corporation
1000, Kamoshida-cho, Aoba-ku, Yokohama, 227, JAPAN
----------------------------------------------------