Re: AW: time per step during GPU-accelerated Implicit Solvent calculations

From: Jim Phillips (jim_at_ks.uiuc.edu)
Date: Mon Nov 16 2015 - 08:55:26 CST

My theory is that due to the lack of periodic boundary conditions your
protein has drifted outside of the initial patch grid and is now entirely
in a single patch so the parallel decomposition is ineffective. If you
restrain the molecule in some way that should limit the performance loss.
Implicit solvent in NAMD was originally intended for use in fitting to
electron densities for which this isn't an issue.

Note that a large ldbperiod won't eliminate the initial load balancing
stage. To completely eliminate load balancing use "ldbalancer none".

Jim

On Mon, 16 Nov 2015, Norman Geist wrote:

> Most likely the NAMD load balancer fails on your side. To verify please add "ldbperiod 100000000" to your namd config and see if this stops the behavior.
>
> Good luck
>
> Norman Geist
>
>
>> -----Urspr√ľngliche Nachricht-----
>> Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im
>> Auftrag von Thomas Albers
>> Gesendet: Montag, 16. November 2015 02:11
>> An: Namd Mailing List <namd-l_at_ks.uiuc.edu>
>> Betreff: namd-l: time per step during GPU-accelerated Implicit Solvent
>> calculations
>>
>> Hello!
>>
>> I am looking at a 200-residue protein using implicit solvent. With 4
>> cores (Xeon E5504) and GPU acceleration (GTX 660), time per step is
>> initially 0.01 s, but after a few 100000 steps climbs to 0.025 steps
>> and more. No other program is taking up CPU cycles, and there's no
>> screensaver or anything that might tax the graphics card. What would
>> cause this strange behaviour, and what should be done?
>>
>> Regards,
>> Thomas
>
>

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:22:14 CST