Project

General

Profile

Task #3220

change rolling pruning scheduling with GPU update

Added by Szilárd Páll 4 months ago. Updated about 1 month ago.

Status:
New
Priority:
Normal
Assignee:
-
Category:
mdrun
Difficulty:
uncategorized
Close

Description

When GPU update is offloaded the list re-pruning can no longer be done "for free" overlapped with CPU tasks, so we should probably just prune in one batch like we do on the CPU.
It might be worth to re-assess whether there is any inter-step overlap to be had (e.g. with other memset or H2D/D2H tasksk) and whether it is worth trading pruning kernel efficiency loss (due to smaller batch execution) with overlapping with other tasks.


Related issues

Related to GROMACS - Feature #2888: CUDA Update and Constraints moduleClosed

History

#1 Updated by Szilárd Páll 4 months ago

  • Related to Feature #2888: CUDA Update and Constraints module added

#2 Updated by Szilárd Páll 3 months ago

  • Target version changed from 2020-rc1 to 2020

Bumped to 2020 final (in case if somebody can pick it up). Only affects performance (and possibly by not much).

Note: need to pass the simulationWorklaod to init_forcerec()->Nbnxm::init_nb_verlet()->setupDynamicPairlistPruning() and set the listParams->numRollingPruningParts = 1.

#3 Updated by Paul Bauer 3 months ago

  • Target version changed from 2020 to 2020.1

I think 2020.1 is a better target

#4 Updated by Paul Bauer about 1 month ago

  • Target version changed from 2020.1 to 2021-infrastructure-stable

more likely for 2021

Also available in: Atom PDF