Hi,
> for a preliminary serial compilation of amber9, a suggestion arose to set
> MKL_HOME to point to intel's mkl mathematical libraries (which I never
> implemented, being limited to libimf.a and libimf.so).
>
> On the other hand
>
> $ ./configure --help
>
> informs that no great improvement can be expected from mkl libraries. I imagine
> the less so for no intel's hardw, such as dual-core-opteron, in my case.
I did 3 (hopefully somewhat) representative benchmarks for sander,
compiled with ifort 10.0.??, with and without MKL (9.0)
What you see in the following are total runtimes of the
bench.jac and bench.cox2 jobs plus a versiob of bench.cox2 without any
cutoff:
core 2 Duo(~2.76 GHz)
MKL noMKL
jac 276.898 312.781
gb_cox2 203.731 218.941
gb_cox2_nocut 1113.535 1477.682
opteron(2.2GHz)
jac 448.207 466.828
gb_cox2 378.946 386.274
gb_cox2_nocut 2433.454 2545.342
My conclusion: Intel (CPU-wise) wins this time and the gain from using the
MKL library is not too great on either machine.
The last time I compared MKL against ACML libraries (with NAB, where the
switch is somewhat straightforward), the AMD libraries lost (=produced
slower programs) even on an opteron.
good luck
Andreas
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Sun Jul 01 2007 - 06:07:23 PDT
