AMBER Archive (2000)

Subject: Re: SGI_MP and sander 6

From: Nathalie Godbout (godbout_at_sgihud.hudson.sgi.com)
Date: Tue Oct 03 2000 - 13:24:52 CDT


Jarrod Smith wrote:
>
> I just noticed that the shared memory parallel option is no longer
> supported in sander version 6. Does using MPI on SGI Origin machines
> introduce a performance hit vs. the old method? For example, if you
> were able to run a calculation on an Origin using both sander classic
> with SGI_MP turned on and using sander 6 with MPI, would you see a
> performance difference?
>
> --
> Jarrod A. Smith
> Research Asst. Professor, Biochemistry
> Asst. Director, Center for Structural Biology
> Computation and Molecular Graphics
> Vanderbilt University
>
> jsmith_at_structbio.vanderbilt.edu

Hi Jarrod,

MPI codes typically perform as well as OpenMP (or sgi $doacross
directives) codes on shared memory systems such as Origin for moderate
number of processors. There is the potential for OpenMP codes to perform
better on shared memory systems because OpenMP offers the ability to do
fine grain parallelism which is difficult to achieve with MPI. But
having said that, it's a difficult question to answer because there are
many factors which can influence the comparison such as the actual
implementation of these parallel paradigms, the size of the calculation,
the type of run etc.

I hope I've answered your questions. I've attached a text file with the
current results I have for the three standard Amber benchmark runs on
sgi platforms running Irix and Linux.

Nathalie

P.S. If you plan on running sander_classic, I've found a bug in one of
the subroutines. Please check-out the following to make the necessary
changes: http://www.amber.ucsf.edu/amber/bugfixes/6.0/bugfix.20

-- 
------------------------------------------------------------------------
Nathalie Godbout, Ph.D.       SGI 
godbout_at_hudson.sgi.com        Chemistry and Biology Applications Group
(781)839-2147                 One Cabot Road, Suite 250
(978)562-7450 fax             Hudson, MA     
vnet:483-2147                 01749, USA
------------------------------------------------------------------------
Amber6/sander (original source code) benchmarks across sgi platforms.

Nathalie Godbout, sgi 09/29/00

*on sgi1400: - 4xPIII 500MHz/2048KB L2 cache - Linux - PGI compilers - MPICH (1.2) and sgi MPI

*on sgi1200: - 16xPIII (8 nodes) 700MHz/256KB L2 cache - Linux - PGI compilers - MPICH (1.2) - ethernet (2 jobs/node) and myrinet (1 job/node)

*on Origin200: - 4xR12k IP27 360MHz/4MB L2 cache - Irix 6.5 - MIPSpro 7.3.1 compilers - sgi MPI

*on sgi2800: - 32xR12k IP27 400MHz/8MB L2 cache - Irix 6.5 - MIPSpro 7.3.1 compilers - sgi MPI

*on sgi3100: - 8xR12k IP35 400MHz/8MB L2 cache - Irix 6.5.9 - MIPSpro 7.3.1 compilers - sgi MPI PROWAT: ------- sgi1400 sgi1400 sgi1200 sgi1200 origin200 (500mhz) (500mhz) (700mhz) (700mhz) (360mhz) mpich sgi-mpi ethernet myrinet ------------------------------------------------------------ nproc time spup time spup time spup time spup time spup 1 278.72 276.80 187.42 186.14 150.21 2 148 1.88 145 1.91 103 1.82 99 1.88 79 1.90 4 87 3.20 84 3.30 64 2.93 55 3.38 45 3.34 8 44 4.26 33 5.64 16 34 5.51 21 8.86 ------------------------------------------------------------ sgi2800 sgi3100 (400mhz) (400mhz) -------------------------------------------- nproc time spup %P time spup %P 1 136.20 133.82 2 71 1.92(95.83) 69 1.94(96.91) 4 39 3.49(95.13) 38 3.52(95.45) 8 22 6.19(95.82) 20 6.69(97.20) 16 13 10.48(96.49) 14 9.56(95.51) 32 11 12.38(94.89) 8 16.73(97.06) --------------------------------------------

DHFR: ----- sgi1400 sgi1400 sgi1200 sgi1200 origin200 (500mhz) (500mhz) (700mhz) (700mhz) (360mhz) mpich sgi-mpi ethernet myrinet ------------------------------------------------------------ nproc time spup time spup time spup time spup time spup 1 441.21 437.73 331.64 330.49 197.67 2 254 1.74 241 1.82 203 1.63 183 1.81 110 1.80 4 157 2.81 146 3.00 146 2.27 101 3.27 61 3.24 8 106 3.13 60 5.51 16 79 4.20 41 8.06 ------------------------------------------------------------ sgi2800 sgi3100 (400mhz) (400mhz) ------------------------------------------- nproc time spup %P time spup %P 1 166.47 163.20 2 94 1.77(87.01) 88 1.85(91.89) 4 51 3.26(92.43) 47 3.47(94.91) 8 30 5.55(93.69) 27 6.04(95.36) 16 20 8.32(93.85) 20 8.16(93.59) 32 18 9.25(92.07) 16 10.20(93.11) -------------------------------------------

4096WAT: -------- sgi1400 sgi1400 sgi1200 sgi1200 origin200 (500mhz) (500mhz) (700mhz) (700mhz) (360mhz) mpich sgi-mpi ethernet myrinet ------------------------------------------------------------ nproc time spup time spup time spup time spup time spup 1 223.14 223.04 166.22 167.95 96.70 2 126 1.77 121 1.84 106 1.57 94 1.79 56 1.73 4 79 2.82 71 3.14 80 2.08 53 3.17 31 3.12 8 56 2.97 33 5.09 16 43 3.86 22 7.63 ------------------------------------------------------------ sgi2800 sgi3100 (400mhz) (400mhz) ------------------------------------------ nproc time spup %P time spup %P 1 85.44 84.18 2 45 1.90(94.74) 45 1.87(93.05) 4 24 3.56(95.88) 24 3.51(95.35) 8 14 6.10(95.55) 14 6.01(95.27) 16 10 8.54(94.18) 10 8.42(94.00) 32 9 9.49(92.35) 7 12.02(94.64) ------------------------------------------