dyer@arktouros.mit.edu (Steve Dyer) (07/24/89)
I'd like to get a feeling for the magnitude of floating point performance increase by adding the AFPA board to a model 125 running the Sep or Dec 88 release of AOS 4.3. I also understand that there are TWO AFPA boards for sale now, one of which was designed for and shipped with the model 125, and one which is shipped with the 130 and 135 (but can be used with the 125, if my understanding is correct.) Is there any reason NOT to buy the newer AFPA board? I believe their prices are similar.
wlm@archet.UUCP (William L. Moran Jr.) (07/24/89)
In article <12926@bloom-beacon.MIT.EDU> dyer@arktouros.mit.edu (Steve Dyer) writes: >I'd like to get a feeling for the magnitude of floating point >performance increase by adding the AFPA board to a model 125 >running the Sep or Dec 88 release of AOS 4.3. > I hate to say this, but my experience is that the eafpa (enhanced advanced ... :) is not a real big win over the regular 125 - mc68881 I've seldom seen performance more than 20% better on pure floating point stuff on a 135 with eafpa vs. 125 with no fpa just a mc881. I suspect this may have something to do with the compiler as I've heard that AIX does much better than this. Anyway, rumors I've heard say that for AOS, the afpa may actually do worse than the mc881, although I don't have an afpa to test this with. Bill Moran -- arpa: moran-william@cs.yale.edu or wlm@ibm.com uucp: uunet!bywater!acheron!archet!wlm or decvax!yale!moran-william ------------------------------------------------------------------------------- Well, I guess if I had to swear one way or another, I'd say he wasn't insane - he just had strange rhythms. Hunter S. Thompson
lmb@ibmpa.UUCP (Larry Breed) (08/09/89)
In article <263@archet.UUCP> wlm@archet.UUCP (William L. Moran Jr.) writes: > >I hate to say this, but my experience is that the eafpa (enhanced >advanced ... :) is not a real big win over the regular 125 - mc68881 >I've seldom seen performance more than 20% better on pure floating >point stuff on a 135 with eafpa vs. 125 with no fpa just a mc881. I >suspect this may have something to do with the compiler as I've heard >that AIX does much better than this. Anyway, rumors I've heard say >that for AOS, the afpa may actually do worse than the mc881, although >I don't have an afpa to test this with. > > Bill Moran You're probably not seeing the best floating point performance AOS offers. The default floating-point code generated by pcc, hc, f77, and hf77 uses the run-time floating linkage (RTFL) so that objects will run on whatever floating hardware is installed. Code tailored for the run-time hardware is generated in data storage the first time a given fp operation is executed. The code is pretty good but can't match the performance of inline code. Only hf77 can generate inline code for the AFPA. Try compiling with the -f2 option, and measure that performance. Larry Breed
grogers@m.cs.uiuc.edu (08/10/89)
I did some simple tests with the whetstone program on several RTs running
AOS (Dec. 1988) with various floating point configurations. Here are the
results.
#define ITERATIONS 100
e_afpa
12.4u 0.1s 0:13 96% 11+23k 0+1io 3pf+0w
afpa
12.8u 0.1s 0:13 94% 12+23k 0+0io 5pf+0w
mc881
16.2u 0.0s 0:16 97% 11+23k 0+0io 3pf+0w
#define ITERATIONS 1000
e_afpa
123.3u 0.2s 2:04 99% 12+23k 0+1io 5pf+0w
afpa
127.2u 0.1s 2:08 99% 12+23k 0+1io 5pf+0w
mc881
161.0u 0.1s 2:41 99% 12+23k 0+1io 7pf+0w
If you're interested, here is where you can get the program that I used.
/* Written 4:09 pm Jul 27, 1989 by davies@uicsrd.csrd.uiuc.edu in m.cs.uiuc.edu:general */
Argonne National Laboratory has an automatic system for fetching
various public-domain software (mostly Fortran numerical code,
but also some in other languages) that includes these and other
benchmarks. The following was returned to me by e-mail by sending
a mail message to:
netlib@anl-mcs.arpa
with the body of the message being:
send index from benchmarks.
To get e.g. whetstone.c, put this in the message:
send whetstonec from benchmarks.
To get general information about netlib, try:
send index.
Jim
==========================================================================
*** from netlib, Tue Feb 14 08:37:11 CST 1989 ***
LINPACKS The LINPACK benchmark program in single precision.
LINPACKD The LINPACK benchmark program in double precision.
LINPACKC The LINPACK benchmark program done in C.
Translated to C by Bonnie Toy 5/88.
LIVERMORE
The Livermore loops in single precision. (Fortran)
A program used by some to benchmark supercomputers.
Frank McMahaon LLNL 1/26/88
LIVERMOREC
A 'C' version of the Livermore loops program,
a program used by some to benchmark supercomputers.
Frank McMahaon LLNL 1/26/88
LUS The linear equation solver used in the paper by Dongarra and Eisenstat
that is based on matrix-vector operations. This is the single
precision version.
LUD The linear equation solver used in the paper by Dongarra and Eisenstat
that is based on matrix-vector operations. This is the double
precision version.
PERFORMANCE
A post script copy of the paper by J. Dongarra, `Performance
of Various Computers Using Standard Linear Algebra Software
in a Fortran Environment'
PERF-UPD
A log of updates made to the above report.
VECTOR a program to check the vectorizing capabilities of a compiler.
D. Callahan, J. Dongarra, & D. Levine, 3/16/88
WHETSTONEC The Whetstone benchmark in C. This program is a translation of the
original Algol version in "A Synthetic Benchmark" by H.J. Curnow
and B.A. Wichman in Computer Journal, Vol 19 #1, February 1976.
WHETSTONES The Whetstone benchmark in single precision Fortran.
WHETSTONED The Whetstone benchmark in double precision Fortran.
DHRYSTONE Benchmark Program in C.
DHRYSTONE-ADA Benchmark Program in Ada.
KARP-CHALLENGE Alan Karp's challenge on parallel vs sequential computing.
NBSLIB A version of netlib runs at NBS and they distribute benchmark
software.
BELL-AWARD details of the following prize:
Beginning in 1988, two $1000 awards will be given each year for 10
years to the person or team that demonstrates the greatest speedup on a
multiple-instruction, multiple-data parallel processor.
NAS A benchmark test program has been developed for use by the NAS program at
NASA Ames Research Center to aid in the evaluation of supercomputer
performance.
David H. Bailey, NASA Ames Research Center, June 24, 1988
NAS-DOC tex documentation for NAS
/* End of text from m.cs.uiuc.edu:general */