Re: ISTAR option and performance

From: Jarrod Smith <jsmith_at_structbio.vanderbilt.edu>
Date: Tue 9 Oct 2001 13:30:22 -0500 (CDT)

You can try this, but be warned it has not been tested to my knowledge and
you will therefore need to verify the results if it works on the SP at
all...

If your compiler/machine/OS are 64bit-clean, you can try setting native=64
in sander.f and compile a large sander that still supports the ISTAR2
scheme. Make sure you tell the compiler to use 64-bit libs, etc. I am
not familiar with the IBM SP series.

I recall compiling this once on SGI as a learning exercise. I left ISTAR2
on, set native=64 in sander.f, and compiled everything with the -64 option
to the MipsPRO compiler. Again, I stress that I did not test the result
so please be careful.

You still might not gain anything depending on how much of a performance
hit you take on switching to 64 bits. But it is something to try (unless
one of the developers knows otherwise).

-- 
Jarrod A. Smith, Ph.D.
Asst. Director, Center for Structural Biology
Research Asst. Professor, Biochemistry
Vanderbilt University
On Tue, 9 Oct 2001, Anton B. Guliaev wrote:
> Dear Amber users,
> 
> we are running sander (Amber 6.0) for rather large systems (over 30000 atoms) at NERSC facility at LBL 
> using IBM SP. We had to recompile sander without ISTAR2 option to accommodate such a large system. 
> After that we observed significant drop in performance (almost 3 times) when we used sander without ISTAR 
> option (4 nodes 16 processors each). Any feedback would be appreciated. Without istar we collected 100000 
> steps in 6.7 hours (200ps). It dropped to 35000 steps using the sander without ISTAR option... thanks a lot -- 
> Anton
>         *******************************
>         *   Anton B. Guliaev, Ph.D    *
>         *   Life Sciences Division    *
>       _ *   LBNL, Berkeley CA 94720   * _   
>      / )*   abguliaev_at_lbl.gov         *( \
>     / / *   F:(510)-486-6488          * \ \
>   _( (_ *   T: 510-486-5773           *  ) )_
>  (((\ \)**/ )*********************( \**(/ /)))
>  (\\\\ \_/ /                       \ \_/ ////)
>   \       /                         \       /
>    \    _/                           \_    /
>    /   /                               \   \
>   /___/                                 \___\
> 

Received on Tue Oct 09 2001 - 11:30:22 PDT