jsh@tut.fi (Helin Jukka) (01/03/89)
Minisupers multi-user systems?
------------------------------
We are buing a Unix-minisupercomputer at Tampere University of Technology
in Finland and we have tested several computers for analyzing the
performance of these machines. The tested machines are Alliant FX/80,
Ardent Titan, Convex C2, ETA-10P, FPS-500 (Celerity 6000) and Stellar
GS1000. Considering the prices and performances for the systems Ardent
Titan seems to be most cost-effective in our computing environment.
However, we (or actually the manufacturer) have run all our these tests
in single-user environment. Does anyone know how how well Titan or
any of these systems support multiple users (both interactive and batch)?
We are looking for a system for 10-30 users (jobs). What about the
reliability of the hardware? And the quality of the software (compilers etc.)?
Because the following performance figures might interest someone, I
have compiled them below. Does anyone have same kind of figures?
A more complete report will be available in short time for anyone who
is interested.
Thanks in advance,
Jukka Helin
Tampere University of Technology
Computing Center
P.O.Box 527
SF-33101 Tampere, FINLAND
tel. +358-31-162388
telefax +358-31-162620
Internet: helin@rapola.tut.fi
Bitnet: HELIN@FINTUTA
uucp: helin@tut.uucp ...mcvax!tut!helin
----------------------------------------------------------------------
We have weighted 11 test programs according to the predicted computing
environment which we will have here, that is:
- 4% program development
- 32% scalar floating point computing
- 60% vector floating point computing
- 4% I/O
Using different weights for test programs we have been able to compile
the following table:
Computer Prog. Scalar Vector I/O Overall
devel. fp fp perf. perf.
perf. perf. perf.
Alliant FX/80 0.33 1.18 3.06 1.63 2.01
Ardent Titan 0.96 1.01 8.14 0.49 3.42
Convex C220 (1 0.89 2.55 15.57 0.79 6.91
Convex C220 (2 0.89 2.55 15.57 11.64 7.69
ETA-10P 0.21 1.02 35.91 0.31 7.74
FPS-500 0.80 2.34 17.14 1.22 7.22
Stellar GS1030 0.48 0.95 6.41 1.96 3.00
(VAX 8700 1.00 1.00 1.00 1.00 1.00 )
1) files created on the disks (I/O tests)
2) files created in the main memory (I/O tests)
Table 1. Program development, scalar floating point, vector
floating point, I/O and overall performances for
single processors.
All the speeds have been normalized to the speed of VAX 8700.
Assuming that there are no bottlenecks in a multiprocessor system,
we can also compute the overall performances for multiprocessor
systems. We simply assume a linear speedup for throughput when
we add more processors. In the following table we have listed some
multiprocessor configurations:
Computer system # of Perf./ Overall
proc. proc. perf.
Convex C240 4 6.9 27.6
Convex C230 3 6.9 20.7
Alliant FX/80 8 2.0 16.0
ETA-10P2 2 7.7 15.4
FPS-500 2 7.2 14.4
Convex C220 2 6.9 13.8
Ardent Titan-4 4 3.4 13.6
Alliant FX/80 6 2.0 12.0
Alliant FX/80 4 2.0 8.0
ETA-10P1 1 7.7 7.7
FPS-500 1 7.2 7.2
Convex C210 1 6.9 6.9
Ardent Titan-2 2 3.4 6.8
Alliant FX/80 2 2.0 4.0
Stellar GS1030 1 3.0 3.0
(VAX 8700 1 1.0 1.0 )
Table 2. Overall performances for some systems.
---