AMBER Archive (2003)

Subject: AMBER: PMEMD 3.01 Update available

From: Robert Duke (rduke_at_email.unc.edu)
Date: Wed Aug 20 2003 - 13:36:56 CDT


Amber Folks -
An update of PMEMD, dubbed 3.01, is now available at the Amber website,
amber.scripps.edu; just click on the PMEMD link. This update mostly covers
support for new machines and a couple of small fixes, detailed below in the
inlined update notice. Otherwise, performance and functionality should be
the same as for 3.00. The copy of the update notice on the web page and
included with the source is more readable than the inlined version. Feel
free to update at your discretion. Any problems, please send me email.

Regards - Bob Duke

Inlined Update Notice:

PMEMD 3.01 (PARTICLE MESH EWALD MOLECULAR DYNAMICS) UPDATE NOTES

This is the update note for a maintenance update of PMEMD 3.00. The major
changes are:
1) Support is added for Compaq or HP AlphaServers, such as those running
at the Pittsburgh Supercomputing Center.
2) Support is added for the Cray T3E.
3) An error in the value printed out for total energy when using NMR
restraints
(nmropt .eq. 1) has been fixed.
4) Amber 7 compatibility mode has been extended in that jfastw is now
handled
in a manner compatible with Sander 7. In 3.00, jfastw was processed in
the manner of Sander 6 regardless of mode, ie., it needed to be set to 1 in
order for changed water names to be picked up and used by the fast analytic
shake routines. Under Sander 7, new water names (watnam, owtnm, hwtnm1,
hwtnm2) are also picked up if jfastw is 0 (default). The consequences of
not having this fix are minor; the "standard" shake routines are used for
water instead of the fast analytic routines, producing insignificantly
different results in slightly more time.
Please read the PMEMD 3.00 release note for the bulk of the details on
PMEMD.
All that is included here is a note on namelists and some benchmarks for the
PSC machines.

NAMELIST ISSUES

Because PMEMD is a f90 program, the standard for namelist input has changed
on
some machines. In the past, the only "problem" platforms have been the IBM
SPx machines, and it is possible to get old format namelists recognized by
including:

For csh:
setenv XLFRTEOPTS "namelist=old"

For Bourne shell:
XLFRTEOPTS="namelist=old"
export XLFRTEOPTS

The Cray T3E, running UNICOS/MK has the same problem, but I am not aware of
an environment variable fix. Thus one needs to change namelist input as
follows:

Old generic format:
Whatever comments you wanted
&namelist_name
foo = bar, ...
&end

New f90-sanctioned format (grrrr...):
!Whatever comments you wanted
&namelist_name
foo = bar, ...
/

Most vendors blessedly recognize the benefits of backward compatibility...

BENCHMARKING RESULTS FOR PITTSBURGH SUPERCOMPUTER CENTER (PSC) MACHINES

Test runs of PMEMD, Sander 6 and Sander 7 have been done on:
lemieux.psc.edu: Compaq 1 GHz ES45 alphaserver, 750x4 cpu, Quadrics
interconnect
salk.psc.edu: HP 1.15 GHz EV7 alphaserver, 1x16 cpu
jaromir.psc.edu: Cray T3E 450 MHz Digital 64 bit alpha chip, 512 cpu
Three protein simulation problems were used. The first is a simulation of a
Factor IX protein with Ca++ ions and a total of 90906 atoms (provided
courtesy
of Dr. Lalith Perera). The second is derived from the JAC benchmark from
Amber 7, a 23528 atom simulation modified in that default values are used
for
cut and skinnb in mdin. The third is identical to the second except that a
constant pressure simulation is done instead of a constant volume
simulation.
All unit cells are orthogonal. Runs were done for 500 steps. An "R" in the
psec/day column indicates that the configuration is "recommended" because at
least 50% scaling is achieved compared to 4 processors on lemieux or
jaromir,
and compared to 1 processor on salk (the higher comparison minimums were
used
because the basic accounting unit is 4 processors on lemieux and because
lower
numbers of processors were not run on jaromir). Runs for Sander 7 can only
be
done on a number of processors that is a power of two, due to a
software-imposed restriction.

****************************************************************************
***
LEMIEUX PERFORMANCE, Compaq 1 GHz ES45 alphaserver, Quadrics interconnect
****************************************************************************
***
90906 Atoms, Constant Pressure Molecular Dynamics (Factor IX)
#procs PMEMD Sander 7 PMEMD/Sander 7
               psec/day psec/day Speedup
  1 48 R
  4 174 R 118 R 1.48x
  8 319 R 186 R 1.72x
12 432 R
16 527 R 240 R 2.20x
24 753 R
32 913 R 299 3.06x
40 1029 R
48 1137 R
56 1223 R
64 1271 322 3.94x
72 1223
80 1271

****************************************************************************
***
SALK PERFORMANCE, 1.15 GHz HP Alpha EV7 SMP
****************************************************************************
***
Runs with more than 8 processors for sander 6 could not be done on salk
due to memory limitations.

90906 Atoms, Constant Pressure Molecular Dynamics (Factor IX)
#procs PMEMD Sander 6 Sander 7 PMEMD/Sander 6 PMEMD/Sander 7
                psec/day psec/day psec/day Speedup
Speedup
  1 62 R 34 R 41 R 1.85x
1.51x
  2 120 R 60 R 99 R 2.00x
1.21x
  4 231 R 126 R 156 R 1.83x
1.48x
  6 339 R
  8 441 R 231 R 310 R 1.91x
1.42x
10 549 R
12 648 R
14 753 R

Runs with more than 8 processors for sander 6 could not be done on salk
due to memory limitations (we probably could have reconfigured sander size
parameters and recompiled, but the trend is obvious).

23558 Atoms, Constant Volume Molecular Dynamics (JAC Benchmark)
#procs PMEMD Sander 6 Sander 7 PMEMD/Sander 6 PMEMD/Sander 7
              psec/day psec/day psec/day Speedup
Speedup
  1 167 R 105 R 131 R 1.59x
1.28x
  2 325 R 179 R 268 R 1.82x
1.21x
  4 626 R 382 R 497 R 1.64x
1.26x
  6 919 R
  8 1200 R 664 R 881 R 1.81x
1.36x
10 1440 R
12 1600 R
14 1878 R

****************************************************************************
***
JAROMIR PERFORMANCE, 450 MHz Cray T3E
****************************************************************************
***
For 90906 atoms, runs on sander 6 and sander 7 were not possible due to
memory limitations (the Cray has only 128 MBytes per processor). In tests a
141,000 atom problem (rt_poly) ran without problems using PMEMD and 8
processors.

90906 Atoms, Constant Pressure Molecular Dynamics (Factor IX)
#procs PMEMD
                psec/day
  4 29 R
  8 56 R
16 108 R
32 195 R
64 334 R
80 393 R
96 393

For 23558 atoms, runs on the copy of sander 7 available on the T3E were not
possible due to memory limitations. I probably could have resized and
recompiled it to get it to run, but didn't.

23558 Atoms, Constant Pressure Molecular Dynamics (JACCP Benchmark)
#procs PMEMD Sander 6 PMEMD/Sander
                psec/day psec/day Speedup
  4 77 R 54 R 1.43x
  8 148 R 99 R 1.49x
16 275 R 171 R 1.61x
32 455 R 268 R 1.70x
64 600~R 360 1.67x
80 635 366 1.73x
96 654 382 1.71x

Bob Duke
University of North Carolina - Chapel Hill
Chemistry Department
rduke_at_email.unc.edu

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu