rosalia@noether.UUCP (Mark Galassi) (05/09/89)
We are debating hotly whether to purchase an SGI 240 with 4
MIPS R3000 processors. One of the attractions is the parallelizing
FORTRAN compiler (yes, we are scientists and there are still some
who program in FORTRAN; sigh).
I came across a report that the parallelizing FORTRAN (what was it,
power FORTRAN or super FORTRAN or whatever) was kind of flaky,
and that the ordinary compilers only worked well without optimization
(this is on the 4 processor machine).
Could someone comment on the soundness of the SGI 4D/240 for number
crunching purposes? We are only slightly interested in the graphics
capabilities: we want a cruncher first of all.
--
{These opinions are mine, and should be everybody else's :-)}
Mark Galassi
rosalia@mozart.UUCP rosalia@sunysbnp.BITNET
rosalia@noether.UUCP rosalia@noether.sunysb.edu bron@bronze.SGI.COM (Bron Campbell Nelson) (05/11/89)
In article <238@noether.UUCP>, rosalia@noether.UUCP (Mark Galassi) writes: > I came across a report that the parallelizing FORTRAN (what was it, > power FORTRAN or super FORTRAN or whatever) was kind of flaky, > and that the ordinary compilers only worked well without optimization > (this is on the 4 processor machine). > There was only one posting I know of that had this complaint. He has not yet given us details of just what his problem is (are you listening Tim?). I (and many others) have used the compiler extensively with very little problem. Both the parallelizer and the optimizer work fine. Several hundred thousand lines of code check out. Now, I'm an engineer, not a salesman, so I'm not going to claim that there are absolutely no problems. But I would claim the compiler is as stable and reliable as the offerings of other companies. BTW - we use the optimizer from MIPSco. They seem to have a reputation for good compiler technology. A single 25MHz cpu gets 4MFLOPS on the Linpack benchmark. Multi- processing speed is completely dependent on the particular code. Some code get nothing; some speed up tremendously. On one of the Livermore Kernels we get about 5.8MFLOPS single processor, and over 20MFLOPS on 4 processors. The best thing to do is to call your nearest SGI office and try to arrange to run your code. The best benchmark is your application. Don't take my word for it; check it out for yourself! -- Bron Campbell Nelson bron@sgi.com or possibly ..!ames!sgi!bron These statements are my own, not those of Silicon Graphics.
mccalpin@loligo.cc.fsu.edu (John McCalpin) (05/13/89)
In article <32454@sgi.SGI.COM> bron@bronze.SGI.COM (Bron Nelson) writes: >In article <238@noether.UUCP>, rosalia@noether.UUCP (Mark Galassi) writes: >> I came across a report that the parallelizing FORTRAN (what was it, >> power FORTRAN or super FORTRAN or whatever) was kind of flaky, >> and that the ordinary compilers only worked well without optimization >> (this is on the 4 processor machine). >I (and many others) have used the compiler extensively with very little >problem. Both the parallelizer and the optimizer work fine. Several >hundred thousand lines of code check out. We have been testing out a 4D/120 and have had reasonable results. The static load distribution technique is only going to give good performance in a single-task environment, but that is what the machine appears to be advertised as, so this is not a problem. I have not had any trouble with the optimizer, though some others here have gotten incorrect results out of their codes. Our conclusion is that the machine is a fine parallel processor for single parallelizable jobs. For multiple users, the throughput of the machine is MUCH higher if all the jobs are compiled and run without the multi-processing option. >The best thing to do is to call your nearest SGI office and try to >arrange to run your code. The best benchmark is your application. Don't >take my word for it; check it out for yourself! Yup. >Bron Campbell Nelson >bron@sgi.com or possibly ..!ames!sgi!bron -- John D. McCalpin - Dept of Oceanography - Florida State University mccalpin@masig1.ocean.fsu.edu mccalpin@nu.cs.fsu.edu mccalpin@fsu (BITNET or MFENET) SCRI::MCCALPIN (SPAN)