Dear David,
Thanks for your kind advice. With your help, I finally found the cause of such a slowdown: nmropt=1. After setting nmropt=0, the speed immediately rised to 113 ns/d. However, since I'm performing an ABFE calculation, the position of the ligand has to be restrained during the TI run. Is there any substitute for nmropt=1?
By the way, I've preformed the test you mentioned. The speed was 102.61 ns/d vs 191.93 ns/d, a TI performance of 53% that of a regular run.
Best,
Kevin
________________________________
From: David A Case <david.case.rutgers.edu>
Sent: Friday, May 12, 2023 0:10
To: Kevin Mei <kevinmei2019.outlook.com>; AMBER Mailing List <amber.ambermd.org>
Subject: Re: [AMBER] Poor Performance of Thermodynamic Integration using pmemd.cuda
On Thu, May 11, 2023, Kevin Mei via AMBER wrote:
>
>I'm now using Amber 20 and RTX 2080Ti to simulate a membrane
>protein system. This system contains ~80,000 atoms in total. In
>regular simulations, the speed is about 110 ns/d, which is quite
>acceptable. However, once I turned on the Thermodynamic Integration
>module, the speed drops sharply to only 24 ns/d. This is quite strange,
>since the manual says "Performance of the PME TI pmemd implementation is
>approximately 75% that of a regular PME MD simulation", and this also makes
>sufficient sampling difficult.
>
>Does anyone have a clue to this? Any help is appreciated.
Try this, just to make sure there is not something special about your
particular system: go to $AMBERHOME/test/cuda/gti/complex, and run the
Run.SC_NVT script (perhaps with a somewhat larger value for nstlim). Then
re-run setting icfe=ifsc=0, and compare timings.
If you do see a big slowdown, this will provide those who really understand
the code a test case to profile. If you don't see that slowdown, we can
start looking for differences between your example and the one in the test
suite. (The next thing I would check is whether setting nmropt=0 makes a
difference.)
....good luck....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu May 11 2023 - 20:00:02 PDT
