-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Hello Swanand--
> Your excellent paper in JACS ("How much backbone motion in ubiquitin...")
> states that "calculation time approximately scales as the inverse of the
> number of CPUs, upto Ne on multiple processor shared memory hardware". Does
> this mean that if 1-member ensemble takes 1 hour, 2-member ensemble is
> expected to take 2 hrs ? or does it mean 0.5 hrs or 1 hour?
That statement simply described the scaling as a function of the
- -num_threads argument: an Ne=2 calculation with -num_threads 1 takes
approximately twice as long as the same calculation with -num_threads 2.
Speed differences between Ne=1 and Ne=2,... calculations are a different
matter. One issue is coupling through the RDC term which can seriously
shorten timesteps chosen by the IVM.
> I have a severe case of slowdown in which single member takes 3 min and
> 2-member takes 1 hr for 1 ps simulation ! This is because the stepsizes in 1
> are much larger than in 2.
so, with -num_threads 1, the calculation takes 60 times longer?
Something is wrong. What are the differences between this calculation ad
the example scripts?
> Then I ran the eginput/gb3_ensemble/tor_cross_two_all.py with ensembleSize 1
> and 2, and got back 10 min and 30 min.
This slowdown of a factor of three is much more along the lines of what
I would expect.
regards--
Charles
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.5 (GNU/Linux)
Comment: Processed by Mailcrypt 3.5.8 <http://mailcrypt.sourceforge.net/>
iD8DBQFFNOjtPK2zrJwS/lYRAj9xAJ9pvTHfaXjFo6uGUPGrVcpZur+viACfUiE/
tPQgiDINBcKZgJUIqBVEyM0=
=CTAL
-----END PGP SIGNATURE-----
