So it looks like a slam dunk to get your test case to at least ~1.5x the speed of the old code. With a little luck, maybe >2x because it's a pretty pathological case for the code that underutilizes the GPU. The new code has the potential to handle that much better by further subdividing work for those instances. I would expect similar results for anything in the 5-15K atom range. Thanks for bringing this to my attention.
Also, simulations that have empty regions of space were initially broken (none of my test cases hit this). That's been addressed.
-----Original Message-----
From: case [mailto:case.biomaps.rutgers.edu]
Sent: Tuesday, February 15, 2011 09:17
To: AMBER Mailing List
Cc: Scott Le Grand
Subject: cuda_DPDP problems; large cutoff radius problems
On Tue, Feb 15, 2011, Scott Le Grand wrote:
>
> For me to fix things at the GPU end if you go that route, I need an
> *explicit* repro with everything explictly set in a single GPU. It'll
> take all of 20 minutes to figure it out with such a simulation (just
> checked in the fix to a repro sent to me at 11:30 last night for
> example).
Scott:
Just checking to see if you saw my earlier posts, with what should be easily
reproducible errors:
http://dev-archive.ambermd.org/201102/0027.html
http://dev-archive.ambermd.org/201102/0028.html
...thx...dac
-----------------------------------------------------------------------------------
This email message is for the sole use of the intended recipient(s) and may contain
confidential information. Any unauthorized review, use, disclosure or distribution
is prohibited. If you are not the intended recipient, please contact the sender by
reply email and destroy all copies of the original message.
-----------------------------------------------------------------------------------
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Feb 17 2011 - 08:30:07 PST
