Performance of IBM 360/91
James MacKinnon
that I used a 360/91 (with a mere two megabytes of memory) while I
was a graduate student at Princeton in the early 70's. I still have a
program that dates from my graduate student days. Here are some timings:
IBM RS/6000 550 (FORTRAN 2.2): 0.07 seconds
IBM 3081/G (Fortran 1.3): 0.25 seconds
Sparcstation 1+ (FORTRAN 1.4): 0.38 seconds
486/25 (Lahey F77L-EM/32): 0.48 seconds
IBM 360/91 (Fortran H): 0.65 seconds
386/33 with 387 (Lahey F77L-EM/32): 0.90 seconds
The 360/91 is faster than any 386 but slower than a 486/25 for this
code, which uses quite a lot of floating point but also involves lots of
subroutine calls and integer calculations. The code is not
matrix-oriented, as can be seen from the fact that the RS/6000 550 is
only 7 times as fast as a 486/25. If it had been, the 360/91 would
probably have done better.
I don't have any record of the timings, but for a while Princeton had a
370/158 as well as the 360/91. The general rule of thumb was that the
360/91 was three times as fast as the 370/158. However, because the
architecture was very different, the 360/91 did a lot better for
floating-point intensive calculations and relatively less well for
integer work.
My conclusion is that, while you might be able to find integer-oriented
programs that don't run any faster on a 360/91 than on a 25Mh 68020 or 386,
most floating-point intensive programs would run a good deal faster on
the 360/91.
Stanley T.H. Chow
>John Mashey's recent post on the performance of the 360/91 reminded me
>that I used a 360/91 (with a mere two megabytes of memory) while I
>was a graduate student at Princeton in the early 70's. I still have a
>program that dates from my graduate student days. Here are some timings:
>
>IBM RS/6000 550 (FORTRAN 2.2): 0.07 seconds
>IBM 3081/G (Fortran 1.3): 0.25 seconds
>Sparcstation 1+ (FORTRAN 1.4): 0.38 seconds
>486/25 (Lahey F77L-EM/32): 0.48 seconds
>IBM 360/91 (Fortran H): 0.65 seconds
>386/33 with 387 (Lahey F77L-EM/32): 0.90 seconds
Considering that the 91 had a 60ns clock with 780 ns memory (from John
Mashey's post), it does exceedingly well. E.g., 386/33 is 30ns clock
usually with <150ns memory.
>The 360/91 is faster than any 386 but slower than a 486/25 for this
>code, which uses quite a lot of floating point but also involves lots of
>subroutine calls and integer calculations. The code is not
>matrix-oriented, as can be seen from the fact that the RS/6000 550 is
>only 7 times as fast as a 486/25. If it had been, the 360/91 would
>probably have done better.
Any more info on the size of the program?
--
Stanley Chow InterNet: sc...@BNR.CA
Bell Northern Research UUCP: ..!uunet!bnrgate!bqneh3!schow
(613) 763-2831
Me? Represent other people? Don't make them laugh so hard.