Re: [AMBER] performance of pmemd.cuda.MPI vs pmemd.cuda both running on single GPU

From: <pavel.banas.upol.cz>
Date: Tue, 16 Oct 2012 22:57:04 +0200 (CEST)

Dear all,

thank you very much for all your answers, especially to Niel for numbers and
Jason for explanation. The reason why I observed worse performance hit was
most likely the size of the system (I used small peptide for testing, ~4500
atoms). Here are numbers I got (also NVT, dt=2fs, ntt=1 and high ntpr, ntwx
and ntwr)



GTX580
pmemd.cuda 128 ns/day
pmemd.cuda.MPI (http://pmemd.cuda.mpi/); 75 ns/day



GTX690
pmemd.cuda 136 ns/day
pmemd.cuda.MPI (http://pmemd.cuda.mpi/); 75 ns/day


Especially when I saw numbers for GTX690 it was for me hard to believe that
it is due to MPI overhead. Now I see that for larger systems the performace
hit become smaller. Thank you very much again for clarification,

Pavel

-- 
Pavel Banáš
pavel.banas.upol.cz
---------- Původní zpráva ----------
Od: Niel Henriksen <niel.henriksen.utah.edu>
Datum: 16. 10. 2012
Předmět: Re: [AMBER] performance of pmemd.cuda.MPI(http://pmemd.cuda.MPI) vs
pmemd.cuda both running on single GPU
"As promised, numbers ....
System: Small RNA, TIP3P water, 7622 atoms
Input: ntt=3, ntp=0, dt=0.002, (NVT)
GPU: Tesla M2090, Keeneland supercomputer
pmemd.cuda, conventional MD: 107 ns/day
pmemd.cuda.MPI(http://pmemd.cuda.MPI), REMD, 24 replicas, exchange attempt 
every 1 ps: 85 ns/day
So yes, there is a performance hit. But not 50%, in this case.
--Niel
________________________________________
From: Niel Henriksen [niel.henriksen.utah.edu]
Sent: Monday, October 15, 2012 1:58 PM
To: AMBER Mailing List
Subject: Re: [AMBER] performance of pmemd.cuda.MPI(http://pmemd.cuda.MPI) vs
pmemd.cuda both running on single GPU
> I'm guessing if Niel compared
> pmemd.cuda to pmemd.cuda.MPI(http://pmemd.cuda.MPI) on the same GPUs (e.
g., on Keeneland) in which
> pmemd.cuda.MPI(http://pmemd.cuda.MPI) ran 1 GPU per replica, he would see 
a performance hit as
> well.
Well, "scheduling is paused" at the moment, but when the job runs I'll
post the numbers ... :)
--Niel
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
(http://lists.ambermd.org/mailman/listinfo/amber)
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
(http://lists.ambermd.org/mailman/listinfo/amber)"
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber

Received on Tue Oct 16 2012 - 14:00:03 PDT