Re: [3.0alpha5] Low global CUDA exclusion count. System is stable without cudaSOAIntegrate or without rigidBonds.

From: Michael Von Domaros (mvondoma_at_uci.edu)
Date: Fri Jul 31 2020 - 11:01:46 CDT

Thanks to both of you for offering to help!

I uploaded all relevant files to this github repository:
https://github.com/mvondomaros/namd3_test

Cheers,
Michael

On Fri, Jul 31, 2020 at 5:01 AM De Carvalho Maia, Julio Daniel <
jmaia_at_illinois.edu> wrote:

> Hi Michael,
>
> Thanks for trying the offloaded integrator.
> If you can send me your files, I can quickly take a look at what’s wrong
> and get back to you ASAP.
>
> Best,
>
> ------------------------------
> *From:* owner-namd-l_at_ks.uiuc.edu <owner-namd-l_at_ks.uiuc.edu> on behalf of
> Michael Von Domaros <mvondoma_at_uci.edu>
> *Sent:* Thursday, July 30, 2020 6:43:15 PM
> *To:* namd-l_at_ks.uiuc.edu <namd-l_at_ks.uiuc.edu>
> *Subject:* namd-l: [3.0alpha5] Low global CUDA exclusion count. System is
> stable without cudaSOAIntegrate or without rigidBonds.
>
>
> Hi everyone,
>
> I'm simulating a single TIP3P water molecule in some lipids (PME, NPT, ambient conditions). The system has been running stable on CPUs for over 20 ns with NAMD 2.14b2. It is also running with NAMD 3.0alpha5 on a GPU if I use the old offloading scheme. Finally, it continues to run well with these versions/configs.
>
> Once I turn on cudaSOAIntegrate, set margin to 4, and remove the pair list updating settings, as recommended, I immediately get:
> FATAL ERROR: Low global CUDA exclusion count! (136304 vs 136307) System unstable or pairlistdist or cutoff too small.
>
> If I turn off rigidBonds, everything is fine again.
>
> Exactly the same system, but with small organic molecules instead of water, is running stable with 3.0alpha5 and with cudaSOAIntegrate + rigidBonds on.
>
> So I thought something might be broken with water, but I can simulate both the apoa1 benchmark and an isolated water molecule (with NAMD 3.0alpha5, cudaSOAIntegrate + rigidBonds on).
>
> I tried increasing pairListDist to 16, reducing the timestep to 0.1 (and at the same time using equal timesteps for all forces), increasing margin to 8, all without success. Visual inspection of the starting structure did not reveal anything odd to me.
>
> Does anyone have suggestions on what else I could try?
>
> Thanks a bunch for your help,
> Michael
>
> I uploaded the config and output files to:
> https://gist.github.com/mvondomaros/0430f2ed7a8409522e3f1f49e897766e
>
>

This archive was generated by hypermail 2.1.6 : Fri Dec 31 2021 - 23:17:09 CST