I'm helping a friend of mine using FSL 4.1.0
I'm not interested in fMRI but I use Linux starting from many years ago.
My friend don't know anything about Linux but he knows about fMRI and so ...
I help her :-)
Really she is interested in tractography.
I've made some benchmark comparison with a standard self-made PC using
Ubuntu Hardy 8.04.1 32bit and 64bit and enabling or disabling SGE for
finding the performance winner.
CPU is a dual core INTEL E6600 Core 2 Duo 2.40 GHz (4MB L2 cache, 1066MHz
FSB, EM64T), total of 2 processors activated (9583.06 BogoMIPS)
Physical RAM memory 2.0 GB DDR2 (2 x 1GB dual-channel)
No memory swapping occours during tests.
Fresh standard 32bit installation without any tricks:
UBUNTU Hardy Heron 8.04.1 Desktop i386 (32 bit i686), updated to today
FSL 4.1.0+fslio1-1~hardy.apsy1 (repository "deb
http://apsy.gse.uni-magdeburg.de/debian hardy main non-free" )
SGE - Sun Grid Engine ver. 6.1u5 (ge-6.1u5-common.tar.gz +
ge-6.1u5-bin-lx24-x86.tar.gz)
/etc/init.d/powernowd stopped (CPU now is in 'performance' mode): leave it
enabled causes problem with SGE (CPU still at 1.60 GHz also if I disable
"ignore_nice_load" in powernowd)
SGE is disabled
Time ('real') "time -p fsl-selftest -c" : 2722 seconds
---------------------
Fresh standard 64bit installation without any tricks:
UBUNTU Hardy Heron 8.04.1 Desktop amd64 (64 bit), updated to today
FSL 4.1.0+fslio1-1~hardy.apsy1 (repository "deb
http://apsy.gse.uni-magdeburg.de/debian hardy main non-free" )
SGE - Sun Grid Engine ver. 6.1u5 (ge-6.1u5-common.tar.gz +
ge-6.1u5-bin-lx24-amd64.tar.gz)
/etc/init.d/powernowd stopped (CPU now is in 'performance' mode): leave it
enabled causes problem with SGE (CPU still at 1.60 GHz also if I disable
"ignore_nice_load" in powernowd)
SGE is disabled
Time ('real') "time -p fsl-selftest -c" : 1959 seconds
------------------------
Ubuntu 64bit seems to be the winner.
And now a test with bedpostx.
My friend has made a 4 slice brain example for testing bedpostx with SGE
Improvement is 89.5% using two core:
** UBUNTU 32 bit **
SGE disabled:
bedpostx /home/alex/FSL/fdt_test_4slice
elapsed 70m56
1 core 100% load
note: 4 slices example = 1 slice processed each time
SGE enabled:
bedpostx /home/alex/FSL/fdt_test_4slice
elapsed 37m25
2 core 100% load
note : 4 slices example = 2 slices processed each time (2 core used)
** UBUNTU 64 bit **
SGE disabled:
bedpostx /home/alex/FSL/fdt_test_4slice
elapsed 133m
1 core 100% load
note: 4 slices example = 1 slice processed each time
SGE enabled:
bedpostx /home/alex/FSL/fdt_test_4slice
elapsed 69m
2 core 100% load
note : 4 slices example = 2 slices processed each time (2 core used)
I'm a bit confused.
fsl-selftest run faster with 64bit.
I've read in the past some benchmark reports about battle Linux 32bit vs. 64bit
All reports say that 32bit works slighty better than 64bit (about 1-10% in
floating point math).
Only if it's used massive integer math, 64bit is slighty better than 32bit.
But bedpostx using ours fdt_test_4slice report that FSL 32bit works MORE
better than 64bit. Double speed !
Is it possible ? Why ?
Cheers
|