Re: [AMBER] GPU Code Update

From: filip fratev <filipfratev.yahoo.com>
Date: Sun, 9 Jan 2011 16:26:56 -0800 (PST)

Hi all,
I can confirm that bugfix.12 seems to work and solved the the issue with NVIDIA GTX4XX cards!!!

I ran several short simulations (200ps each) using JAC-PME cuda benchmark and GTX470 card without any problems! Previously this was just impossible. The performance (ns/days) was identical to my last tests, but for GB it works better now - TRPCage 398 ns/day against 380 ns/days before.
Thank you for this bugfix!

Regards,
Filip
         

--- On Sat, 1/8/11, Ross Walker <ross.rosswalker.co.uk> wrote:

> From: Ross Walker <ross.rosswalker.co.uk>
> Subject: [AMBER] GPU Code Update
> To: "AMBER Mailing List" <amber.ambermd.org>
> Cc: "Duncan Poole" <dpoole.nvidia.com>, "Scott Le Grand" <SLeGrand.nvidia.com>, "Slaton Lipscomb" <lipscomb.slaton.gene.com>
> Date: Saturday, January 8, 2011, 6:57 PM
> To all those who are waiting for the
> latest update to the GPU code,
>
> Firstly I apologize for the day here but New Years of
> course bring new
> deadlines but I have now managed to put this together and
> it is posted on
> the AMBER website as bugfix.12. I would suggest that
> instead of just
> applying this bugfix you instead start from a completely
> new Amber 11 tree
> as follows:
>
> tar xvjf AmberTools-1.4.tar.bz2
> tar xvjf Amber11.tar.bz2
> export AMBERHOME=/path_to_/amber11
> cd $AMBERHOME
> wget http://ambermd.org/bugfixes/AmberTools/1.4/bugfix.all
> patch -p0 <bugfix.all
>
> rm -f bugfix.all
> wget http://ambermd.org/bugfixes/11.0/bugfix.all
> wget http://ambermd.org/bugfixes/11.0/apply_bugfix.x
> chmod 700 apply_bugfix.x
> ./apply_bugfix.x bugfix.all
>
> Then just build things again from scratch.
>
> The specifics of the bugfix are as follows:
>
> ********>Bugfix 12:
> Author: Ross Walker
> Date: 20 December 2010
>
> Program(s): pmemd.cuda
>
> Description: - This fixes a number of recently discovered
> bugs and possible
>            
>    performance issues with pmemd.cuda.
> Specifically it does the
>            
>    following:
>
>            
>    1) Removes the dependence on CUDPP which
> could cause
>                
>   crashes on certain large simulations due to bugs in
>                
>   the CUDPP library.
>
>            
>    2) Fixes problems with large simulations,
> 400K+ atoms
>                
>   crashing randomly with an allocation failure.
>
>            
>    3) Fixes "ERROR:   max
> pairlist cutoff must be less than unit
>                
>   cell max sphere radius!" bug allowing cutoffs larger
> than
>                
>   8.0 to be used for both small and large systems.
>
>            
>    4) Writes final performance info the the
> mdout file.
>
>            
>    5) Possible (not fully tested) workaround
> for NVIDIA GTX4XX
>                
>   and GTX5XX series cards to avoid possible random
> hangs
>                
>   during PME simulations.
>
>            
>    6) Tests for the use of non-power of 2 MPI
> tasks when running
>                
>   parallel GPU PME calculations. Update so that the
> code now
>                
>   quits with a suitable error message.
>
>            
>    7) Minor performance improvements, mostly
> due to the use of a
>
>                
>   faster radixsort.
>
> Apply this patch in $AMBERHOME
>
>
> All the best
> Ross
>
> /\
> \/
> |\oss Walker
>
> ---------------------------------------------------------
> |         
>    Assistant Research Professor   
>           |
> |            San Diego
> Supercomputer Center         
>    |
> |         
>    Adjunct Assistant Professor   
>            |
> |         Dept. of Chemistry
> and Biochemistry       
>    |
> |          University of
> California San Diego       
>    |
> |               
>      NVIDIA Fellow     
>            
>    |
> | http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
> | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk 
> |
> ---------------------------------------------------------
>
> Note: Electronic Mail is not secure, has no guarantee of
> delivery, may not
> be read every day, and should not be used for urgent or
> sensitive issues. 
>
>
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>


      

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Jan 09 2011 - 16:30:02 PST
Custom Search