[AMBER] Benchmarking sander.MPI and pmemd on a linux cluster with infiniband switch

From: Ilyas Yildirim <i-yildirim.northwestern.edu>
Date: Tue, 11 Jan 2011 14:58:54 -0600 (CST)

Hi All,

I am benchmarking 3 systems on a linux cluster with infiniband switch
before submitted my jobs. I have compiled amber9 and pmemd using
intel/11.1-064 compilers, and mpi/openmpi-intel-1.3.3.

There are 2 types of nodes in the system, which I am benchmarking.

i. 8 core nodes (Intel Xeon E5520 2.27GHz - Intel Nehalem) - old system
ii. 12 core nodes (Intel Xeon X5650 2.67GHz - Intel Westmere) - new system

The 3 systems have 63401, 70317, and 31365 atoms, respectively. Here are
the results:

###########################################################
# System # 1:
# 63401 atoms (62 residues, 540 Na+/Cl-, 60831 WAT)
#
# old Quest
# (Intel(R) Xeon(R) CPU E5520 . 2.27GHz - 8 core/node)
#
# pmemd (hrs) sander.MPI (hrs)
8 1.32 1.78
16 0.77 1.28
24 0.64 1.02
32 0.50 0.95
40 0.44 0.88
48 0.41 0.87
56 0.41 0.87
64 0.40 0.85
72 0.39 0.85
80 0.39 0.87
#
# new Quest
# (Intel(R) Xeon(R) CPU X5650 . 2.67GHz - 12 core/node)
#
# pmemd (hrs) sander.MPI (hrs)
12 0.86 1.23
24 0.55 0.94
36 0.41 0.82
48 0.36 0.82
60 0.32 0.75
72 0.32 0.77
84 0.31 0.73
96 0.31 0.78
#
###########################################################

###########################################################
# System # 2:
# 70317 atoms (128 residues, 1328 Na+/Cl-, 64689 WAT)
#
# old Quest
# (Intel(R) Xeon(R) CPU E5520 . 2.27GHz - 8 core/node)
#
# pmemd (hrs)
8 1.35
16 0.81
24 0.62
32 0.51
40 0.46
48 0.43
56 0.41
64 0.42
72 0.40
80 0.39
#
# new Quest
# (Intel(R) Xeon(R) CPU X5650 . 2.67GHz - 12 core/node)
#
# pmemd (hrs)
12 0.89
24 0.56
36 0.43
48 0.37
60 0.33
72 0.32
84 0.32
96 0.31
#
###########################################################

###########################################################
# System # 3:
# 31365 (28 residues, 680 Na+/Cl-, 26382 WAT, 3430 heavy atom)
#
# (hours)
# sander.MPI sander.MPI(new)
8 0.91 0.91
16 0.63 0.63
24 0.55 0.54
32 0.52 0.52
40 0.49 0.49
48 0.50 0.50
56 0.50 0.50
64 0.53 0.53
72 0.47 0.46
80 0.47 0.47
#
# new Quest
# (Intel(R) Xeon(R) CPU X5650 . 2.67GHz - 12 core/node)
#
# sander.MPI (hrs)
12 0.62
24 0.49
36 0.46
48 0.45
60 0.47
72 0.38
84 0.39
96 0.40
#
###########################################################

It seems that I am hitting the peak around 48 cpus. In the amber mailing
list, I found some threads where Ross Walker and Robert Duke discusses the
efficiency and scaling of pmemd. For a system with over 70K, I am unable
to get a peak around 128 cpu, which Ross was talking in one of the thread
(for a system with 90K atoms). Therefore, I have some questions and will
appreciate any comments.

1. How does sander.MPI and pmemd divide the system when multiple cores are
used? Does it divide the system randomly or according to the number of
residues (excluding water and ions)?

2. Is these results compatible with anyone's experience? I heard that with
LAMMPS and NAMD, people can get a good scaling up to 256 cores (for
systems with 1 millions of atoms). Just for curiosity; would pmemd scale
efficiently on a system with over 1 millions of atoms?

3. I am using AMBER9. Does the scaling get better on AMBER10 or AMBER11?

4. In system # 3, I cannot use pmemd because of the positional restraints
imposed on the system. Can I use the new versions of pmemd with positional
restraints?

Thanks in advance. Best regards,

   Ilyas Yildirim, Ph.D.
   -----------------------------------------------------------
   = Department of Chemistry - 2145 Sheridan Road =
   = Northwestern University - Evanston, IL 60208 =
   = Ryan Hall #4035 (Nano Building) - Ph.: (847)467-4986 =
   = http://www.pas.rochester.edu/~yildirim/ =
   -----------------------------------------------------------


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Jan 11 2011 - 13:00:07 PST
Custom Search