pack@acd.ucar.edu (Dan Packman) (12/08/90)
The following very short benchmark program attempts to measure
raw fortran unformatted io speed. I can duplicate the program n times
for n tmp files or can run n read programs on a single tmp file.
The tmp file is about 10 megabytes in size. The 130000 dimension
is representative of some of our applications.
I simply run "simultaneously" via
% time rd & ; time rd1 & ; time rd2 & ; time rd3 & ; ...
Running on an IBM RS6000/320 I obtain (in elapsed seconds on an
otherwise unloaded machine):
\n | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
__________________________________________________________________
one tmpfle | | 31 | 47 | 64 | 80 | 101 | 119 |
_______________16.5_______________________________________________
n tmpfles | | 36.5| 56 | 86-116 |
__________________________________________________________________
The first test shows remarkable linearity through 7 simultaneous processes.
The second test with multiple files to read shows a falloff (and unequal
times for the processes) for 4 simultaneous processes. Any comments?
[I would guess that "server" class machines might show slightly
better times for a single read, but would show much better results
for many simultaneous processes.]
=====
program wr
double precision a(130000)
open(1,file='/tmp/tmpfle1',form='unformatted',status='new')
do 10 i=1,10
a(130000)=dble(i)
write(1)(a(ii),ii=1,130000)
10 continue
end
program rd
double precision a(130000)
open(1,file='/tmp/tmpfle1',form='unformatted',status='old')
do 10 i=1,10
read(1)(a(ii),ii=1,130000)
if (a(130000).ne.dble(i)) then write(0,100)i,a(130000)
100 format(' i=',i2,' bad value =',1pd12.5)
10 continue
end
Dan Packman NCAR INTERNET: pack@ncar.UCAR.EDU
(303) 497-1427 P.O. Box 3000 CSNET: pack@ncar.CSNET
Boulder, CO 80307 DECNET SPAN: 9.367::PACKmark@mips.COM (Mark G. Johnson) (12/11/90)
In article <9456@ncar.ucar.edu>, pack@acd.ucar.edu (Dan Packman) gives a short benchmark program that exercises I/O. He presents measured data for n simultaneous processes reading the same file, and also data for k processes simultaneously reading different files. Dan writes: >The first test shows remarkable linearity through 7 >simultaneous processes. I repeated Dan's test of simultaneous processes reading the same file (using Dan's pgm to create that specific file) and found a modest degree of better-than-linearity (slope a bit less than one). Measurements on the two workstations >in elapsed seconds on an otherwise unloaded machine: # of simultaneous processes reading same file Hardware 1 2 3 4 5 6 7 ---------------------------------------------------------------------------- IBM RS6000/320 [dan] 16.5 31 47 64 80 101 119 MIPS RS3230 Magnum [mgj] 11.3 21.1 31.0 40.9 50.7 60.8 71.0 The (slight) difference between the two machines is the incremental cost of the 2nd process. For the IBM, each process costs about 16.5 seconds. For the Magnum, each process costs around 9.9 seconds, *except* the first one which costs 11.3. Also the IBM begins to depart from linearity at n>5 processes, when the cost per process rises again. At least thru 7 processes the Magnum seems not to do this. -- -- Mark Johnson MIPS Computer Systems, 930 E. Arques M/S 2-02, Sunnyvale, CA 94086 (408) 524-8308 mark@mips.com {or ...!decwrl!mips!mark}