Project

General

Profile

Task #2296

tuning single-rank PME GPU scheduling

Added by Szilárd Páll almost 2 years ago. Updated over 1 year ago.

Status:
Closed
Priority:
Normal
Category:
mdrun
Target version:
Difficulty:
hard
Close

Description

I have identified a few cases that would benefit from tuning for the use-cases the 2017 release will support with PME-GPU.

  1. Placement of 2nd phase of PME (FFT + gather) wrt other tasks

Esepecially with small inputs, the increasing launch overhead of (especially) the FFT kernels can end up delaying the launch of nonbonded tasks to the extent very little to PME - nonbonded task overlap happens. Launching this second stage after the nonbonded tasks are launched improves performance significantly. This can however have a negative impact on the PME performance which needs to be investigated.

  1. Alternating GPU task wait/reduction

Instead of blocking waiting in a statically defined order for one of the GPU tasks, to allow the overlap of the reduction of forces arriving from the GPU task that finishes earlier, we can implement a alternating polling scheme. (WIP)

History

#1 Updated by Szilárd Páll over 1 year ago

  • Private changed from Yes to No

Szilárd Páll wrote:

I have identified a few cases that would benefit from tuning for the use-cases the 2017 release will support with PME-GPU.

  1. Placement of 2nd phase of PME (FFT + gather) wrt other tasks

Done in https://gerrit.gromacs.org/7212

  1. Alternating GPU task wait/reduction

Instead of blocking waiting in a statically defined order for one of the GPU tasks, to allow the overlap of the reduction of forces arriving from the GPU task that finishes earlier, we can implement a alternating polling scheme. (WIP)

Done in https://gerrit.gromacs.org/#/c/7182/

#2 Updated by Szilárd Páll over 1 year ago

  • Status changed from New to Resolved

#3 Updated by Szilárd Páll over 1 year ago

PS: This can be closed, but I'd suggest we keep it open for taking notes on the results of testing the new PME waiting code on different modes/hw. (or open a new task?)

#4 Updated by Mark Abraham over 1 year ago

  • Status changed from Resolved to Closed

new task sounds fine

Also available in: Atom PDF