On Sat, Aug 28, 2010, Ross Walker wrote:
>
> AMBER 11 with and without the fix for FOPTFLAGS in the Makefile.
>
> Original Makefile (PMEMD_OPTFLAGS)
> -----------------
>
> Compiler States Vectorization at link? JAC benchmark time
> (s)
> -------- -----------------------------
> ----------------------
> gfortran 4.1.2 N/A 178.15
> ifort 10.1.018 Yes 155.60
> ifort 11.0.084 Yes 149.33
> ifort 11.1.056 No 149.30
> ifort 11.1.069 No 150.40
What do you mean by "states vectorization at link"? What flags did you
give to configure? And which exact benchmark? Maybe you should state
what sort of chip you have as well.
>
> Dave, what version of ifort were you using and what is the difference in
> timings? Do you see the words:
>
> ipo: remark #11000: performing multi-file optimizations
> ipo: remark #11005: generating object file /tmp/ipo_ifort8APlbU.o
I get the above both before and after the bugfix. But after, the link step
takes a lot longer than before, say a few seconds vs. about a minute.
Running the "bench.jac.pmemd" script, with mpirun -np 8 as my do-parallel,
I get 638 seconds before the bugfix, and 164 seconds after. Ifort 11.1.069,
but on intel nehalem (X5570 . 2.93GHz). Note that 164 sec corresponds
to 10.5 ns/day.
I don't understand diffs between you and me, but the bugfix seems worth
posting...won't do any harm and is what was actually intended.
...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sat Aug 28 2010 - 10:30:03 PDT
