Dear Robert, list,
for the record: Robert was right, setting maxmem to 2^30 (we have 4GB of
RAM) brought the estimation to a speed even below the 30 min spm2 needed
for estimating a similar model, resulting in an 8x speedup of the
estimation process.
It does of course make sense to blame low maxmem settings for low speed
(due to large amounts of disk i/o), but in my tests the setting of
maxmem in spm5 was identical to spm2 (2^20), and yet I did observe those
huge performance differences. Therefore I ruled out that explanation
initially and didnt test it further. I still can't explain the
differences between spm2 and spm5 with identical maxmem settings, but it
does not really matter now spm5 is fast enough ;-)
Still it would be interesting to know whether Xeon 64 bit still suffers
from the NaN problem or not, to even further speed up processing (one
never has enough speed!).
Thanks for the useful advice!
Cheers,
Bas
Robert Welsh wrote:
>You should change maxmem
>
>FYI : 2^20 = 1megabyte.
>
>you are most likely thrashing yourself to death.
>
>if you have a 2 gig of ram i would suggest making
>
>maxmem = 2^28 or 2^29
>
>that is 256meg and 512meg respectively. if you have more memory then
>increase. i suspect that you are spending alot of time reading and
>writing data to disk, hence not maxing out on cpu cycles. sounds like
>you design is quite simple, just one regressor which should be quite
>speedy.
>
>robert welsh
>
>
>-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
>Robert C. Welsh, PhD
>Research Investigator
>Department of Radiology
>University of Michigan
>(734) - 764 - 2412 (fax)
>[log in to unmask]
>
>
>
>
>
>
>>>>Bas Neggers <[log in to unmask]> 01/09/07 6:52 AM >>>
>>>>
>>>>
>Dear list,
>
>we are currently planning to migrate our fMRI postprocessing from spm2
>to spm5. I (a researcher and spm5 user, no sysadmin) did a couple of
>tests, using a very similar and simple model (1 box- car HRF
>convolved
>regressor + constant) in both spm2 and spm5, 1000 scans. What is
>really
>strange to me is that in spm2 the estimation takes about 30 minutes (2
>passes; AR(1) autocorrelation estimation, actual parameter
>estimation),
>whereas in spm5 it took about 4 hours (same 2 passes, output in
>console
>is very similar). Furthermore, CPU load for the spm2 estimation jumped
>to almost 90%, whereras for spm5 it hardly exceeded 20%. That all
>would
>be unworkable with more than 10 users. We use a multiprocessor Xeon
>Linux server, running debian, and run both spm's in matlab 7 (R2006b).
>I did recompile the spm5 (mexa64) binaries on our system, using the
>default spm5 makefile in /src (make; make install). The last spm5
>patches we applied where updates_573.
>
>Preprocessing in spm5 appeared to be of comparable speed as compared
>to
>spm2.
>
>I have set defaults.stats.maxmem to 2^20 in both spm2 and spm5.
>
>I really can't explain this discrepancy. In the archives someone
>seemed
>to have noticed something similar:
>http://www.jiscmail.ac.uk/cgi- bin/wa.exe?A2=ind0610&L=SPM&P=R18235&I=-
>3&X=09940203E3366DA97F&Y=b.neggers%40umcutrecht.nl
>
>Our problem is different from the batch interface handling slowing
>discussed here:
>http://www.jiscmail.ac.uk/cgi- bin/wa.exe?A2=ind0611&L=SPM&P=R37959&I=-
>3&X=09940203E3366DA97F&Y=b.neggers%40umcutrecht.nl
>That seemed to have been caused by file loading in the batch
>structure,
>we do not seem to have that issue.
>
>But that sysadmin was told not to worry since fMRI estimations do take
>a
>lot of time, and "get used to it". But with my tests in mind I am
>afraid
>that something has fundamentally changed in the estimation procedure,
>which slows estimation down, be it perhaps on only some systems. I
>have
>thought about the NaN issue affecting Intel processors, which is
>indeed
>thought to slowdown estimation. But I recompiled the binaries.
>Furthermore, the processor seems not to be taxed to the full in spm5,
>which one would not expect with NaN problems.
>Perhaps spm5 has other file I/O bottlenecks/memory mapping techniques?
>Should I still recompile with other gcc settings than default in spm5?
>On this page:
>
>http://www.sph.sc.edu/comd/rorden/spm99.html
>
>...you can read about slow spm2 processing on intel CPUs, but nothing
>is
>said if this is dealt with correctly by spm5 makefile/c- code. Should
>I
>change the gcc flags in the spm5 makefile too, or do 64 bit Xeon
>processors have no issues anymore with NaN? On the above benchmark
>site
>there is no mention yet of 64 bit procs nor spm5...
>
>Any help would be appreciated!
>
>Cheers,
>
>Bas
>
>--
>--------------------------------------------------
>Dr. S.F.W. Neggers
>Division of Brain Research
>Rudolf Magnus Institute for Neuroscience
>Utrecht University Medical Center
>
>Visiting : Heidelberglaan 100, 3584 CX Utrecht
> Room A.00.1.24
>Mail : Huispost A.01.126, P.O. Box
> 3508 GA Utrecht, the Netherlands
>Tel : +31 (0)30 2503386
>Fax : +31 (0)30 2505443
>E- mail : [log in to unmask]
>Web : http://www.fmri.nl/people/bas.html
>--------------------------------------------------
>
>
>
>**********************************************************
>Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues.
>
>
--
--------------------------------------------------
Dr. S.F.W. Neggers
Division of Brain Research
Rudolf Magnus Institute for Neuroscience Utrecht University Medical Center
Visiting : Heidelberglaan 100, 3584 CX Utrecht
Room A.00.1.24
Mail : Huispost A.01.126, P.O. Box
3508 GA Utrecht, the Netherlands
Tel : +31 (0)30 2503386
Fax : +31 (0)30 2505443
E-mail : [log in to unmask]
Web : http://www.fmri.nl/people/bas.html
--------------------------------------------------
|