Hi Bernd,
You're quite correct about the graphics performance within SPM, though
it's probably a useful thing to know if your rendering software uses
OpenGL to display brain activity e.g. BrainX.
The influence of running SPM across the network (and its impact on
graphics performance) can best be seen in comparing the results from a
SUN with 8*750MHz UltraSPARC III CPUs, see Matlab 6.x results D & M
(over network, on console respectively)
I hadn't really intended to comment on memory bandwidth - though I
suspect that CPU speed will be the overriding influence on total speed.
Though there is an interesting result for a 2GHz P4 (Linux, with 2Gb of
RAMBUS DRAM) coming out slower than a 1.2GHz Athlon (Win98, 512Mb
DDR-RAM(?)) .....
I totally agree that the I/O bandwidth of disk access wasn't addressed
in this simple test and is probably quite important. Though what
proportion of time a machine spends swapping memory to/from disk will
depend on the amount of physical RAM installed, the size of the dataset
etc etc. I.e. difficult to quantify.
As for the jvm versus nojvm options look at K/L and Q/R (jvm/nojvm) on
the matlab 6.x results, there doesn't seem to be much in it.
Yours,
Jon.
Jonathan Brooks (Ph.D.) - Research Fellow
MARIARC, University of Liverpool, Liverpool, UK
Tel: +44 151 794 5629 Fax: +44 151 794 5635
Web: http://www.mariarc.liv.ac.uk
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Bernd Markgraf
Sent: 04 April 2002 16:48
To: [log in to unmask]
Subject: Re: Benchmarking (at last)
hi jon,
thanks for your compilation of matlab benchmark results. just some
thoughts about the whole thing:
1. the benchmark includes 2d and 3d test which in my case gives very bad
results since i'm running matlab remotely on the server. since
graphic
performance isn't all that important for matlab i would suggest to
ignore the last two test.
2. matlabs own benchmark doesn't say much about the performance of the
whole machine since basically all tests run inside cpu cache. so
memory bandwidth is completely ignored. the same is true for disk i/o
which is not tested.
3. for matlab 6 it would be interesting if the benchmark was run with or
without jvm.
bernd
|