| Some benchmarks [message #4026] |
Wed, 19 April 1995 00:00  |
uphlabh
Messages: 8
Registered: December 1994
|
Junior Member |
|
|
I noted the hoopla about benchmarks running on pentiums, SGI's, suns,
and about every other machine, and I thought I would include some local
machines for this list. This is running the TIME_TEST benchmark.
SGI 4D/340S (4 CPU) 37.2 seconds {usual loaded-down state}
4D/310S (1 CPU) 39.9 {loaded}
SGI Indy 17.7 {nobody}
DECstation 5000/200 79.4 {nobody on}
DEC Alpha (64 MB RAM) 13.3 {two other jobs}
DEC Alpha (254 MB) 7.0 {nobody}
The Alpha times seem to indicate this who procedure is heavily
swap-dependent, which is not too big a surprise I suppose. The big
SGI's aren't too revealing, as they are pretty loaded down and I didn't
ever get a chance to try them otherwise.
Regards,
Brian
|
|
|
|
| Re: Some benchmarks [message #4076 is a reply to message #4026] |
Tue, 25 April 1995 00:00  |
clodius
Messages: 2
Registered: April 1995
|
Junior Member |
|
|
In Article <gurman-1904952039240001@arrowroot.gsfc.nasa.gov>,
gurman@uvsp.gsfc.nasa.gov (Joseph B. Gurman) wrote:
> <snip>
> For what it's worth, with an unreleased beta version of IDL for the Power
> Macintosh, we get:
>
> PowerMac 8100/80 48 Mbyte System 7.5.1 20.9 s
>
> Frankly, I think disk speed is a big factor in these, as the Power Mac figures
> (test 23 = 7.38 s) shows in comparison to say, the 3000/600 running OSF/1
> (test 23 = 0.82 s). The OpenVMS systems also have extra overhead in disk
> writing (test 23 = 2.53 s on the 3000/900, 6.95 s on the 4000/710), but
> you supposedly get a more reliable file system in return. In fairness, the
> PowerMac disk I/O is still done in emulation.
>
> Joe
>
> P.S. Bet our PowerMac cost less than your Indy, though.... It certainly
> cost about a factor of 3 less than the DEC 3000/400.
>
> --
> Joseph B. Gurman / NASA Goddard Space Flight Center/ Solar Data Analysis Center
/ Code 682 / Greenbelt MD 20771 USA / gurman@uvsp.gsfc.nasa.gov
> | Federal employees are still prohibited from holding opinions while at work.
Any opinions expressed herein must therefore be someone else's. |
Be aware that time_test for the Power Macs, and probably for a lot of other
processors is dependent on more than just processor, raw memory, clock
speed, and disk speed. I had an opportunity to try out IDL in a class on a
Power Mac 6100 with 16 Mbytes of memory. Experimenting with time_test
showed that it also depended on whether
1. It was run the first or latter time. On the first time you had
additional disk accesses and compilation to deal with so that the second run
was significantly faster.
2. The disk cache and program had adequate memory set aside (critical on a
16 Mbyte machine as the version of IDL reccomends 10 Mbytes + for IDL
alone). If the cache was large enough, disk access was less of a problem.
3. Graphics were written to a buffer for updating. Without the buffer you
had some problems with the appearance of windows ;-), but about twice the
nominal graphics performance.
4. Modern Memory manager and Virtual memory were used.
5. Sometimes performance would change with configuration in a nonobvious
way, resulting in a factor of four or more degradation in performance.
6. Whether the beta or older non beta version was used. The beta version
was typically more stable and faster than the nonbeta version, but could not
run a few of the files.
For the record, with graphics buffering, a beta version of IDL on the 6100
Power PC, for the first run of time test I remember getting results between
30 and 70 s, for the second run between 17 and 70 s. The non-beta version
was 25-50% slower. A 90 MHz Pentium nearby was reporting about 16 s, and
over twice the graphics performance. Most of the performance was determined
by the disk intensive routines, which varied by over an order of magnitude
depending on the cache, etc. settings. I recall test 23 taking 3-4 seconds
with the optimum setting on the second run, and more than 20 seconds on the
worst setting.
In the class itself, where routines were applied to arrays that could never
fit in the disk cache on a 16 Mbyte machine, performance was at best a
factor of two worse than the Suns and PCs. It was also significantly less
stable than the Suns with most of the problems due to memory conditions, but
also had problems with directories on CD-ROMs where it would try to perform
a write operation ;-) andd bomb instead of failing gracefully. It appeared
to also be less stable than the PCs, but the PCs also bombed a number of times.
I do not recomend running IDL on a 16 Mbyte Power Mac, 24 Mbytes may be
adequate, but you can never have too much memory.
------------------------------------------------------------ ------------
William B. Clodius, wclodius@lanl.gov
NIS-1, Los Alamos National Laboratory
------------------------------------------------------------ ------------
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
