Re: QMMM latency in PME correction

From: Marcelo C. R. Melo (melomcr_at_gmail.com)
Date: Wed Nov 17 2021 - 21:01:19 CST

Hi Michael,
Thank you for sharing the issue here.

To follow up with some specifics, the issue is arising in a system with
approximately five thousand QM atoms and 12 thousand point charges (PCs)
representing MM atoms.
The PME correction is currently executed in a single core and depends on
two double loops, one that treats all QM-QM atom pairs and another that
treats QM-PC pairs.

My thought is that the PME correction could be distributed over all ranks
that have relevant QM and/or MM atoms by sharing the positions and charges
of QM atoms. Ths would require a couple of extra messages for each QM step,
which would not be a problem for a large system like this, but I wonder
what that would entail for smaller QM regions.

We would welcome any thoughts on this. In the meantime, I'll run some extra
timings and benchmarks to see if this approach is feasible.

Best,
Marcelo

On Tue, 16 Nov 2021 at 16:41, Wall, Michael E <mewall_at_lanl.gov> wrote:

> Hello NAMD-L,
>
> I'm using NAMD to perform QMMM using a large QM region, using LATTE as a
> QM solver (https://urldefense.com/v3/__https://github.com/lanl/latte__;!!DZ3fjg!qbNSrKdN68fL0G2-A-4O-3yn-dd6Pna-RbcGZql0DnmUTI8BFhidjtagkbdT2TV6mQ$
> <https://urldefense.com/v3/__https://github.com/lanl/latte__;!!DZ3fjg!sF2-qGbr94sZuQylawLrZXrTe3zqvOL5SUo1Po3pBtX5gf_92TKt2iTMDez8dOCyhg$>).
> I'm seeing some latency on the NAMD side that I'm trying to address.
>
> I've done some profiling and found that most of the latency is associated
> with a PME correction in src/ComputeQM.C (line 5033 in the NAMD_2.14b2_Source
> version, about 6-9 seconds required for the system I'm currently looking
> at).
>
> I had a conversation with Marcelo Melo about strategies for
> addressing this issue by parallelizing the PME correction (along with a
> couple of other calculations). He shared his thoughts and also suggested
> I post to this mailing list to share the observation more broadly and
> to see whether one of the other developers might have any insights.
>
> Thank you for considering this issue.
>
> Sincerely,
>
> Michael Wall
>
>
> ---
> Michael Wall
> mewall_at_lanl.gov
> https://urldefense.com/v3/__http://public.lanl.gov/mewall__;!!DZ3fjg!qbNSrKdN68fL0G2-A-4O-3yn-dd6Pna-RbcGZql0DnmUTI8BFhidjtagkbeO1dWyTw$
> <https://urldefense.com/v3/__http://public.lanl.gov/mewall__;!!DZ3fjg!sF2-qGbr94sZuQylawLrZXrTe3zqvOL5SUo1Po3pBtX5gf_92TKt2iTMDeyDZi9pJQ$>
>
>

This archive was generated by hypermail 2.1.6 : Fri Dec 31 2021 - 23:17:12 CST