Getting best performance from GPUs while using Quantum ESPRESSO

Accelerated Computing GPU-Accelerated Libraries
1

Dear all,
I am trying to use Quantum ESPRESSO with GPU acceleration.
However, since I am new to this field, I do not know about the flags and variables that could be exported to get the optimum performance.

Here is a bunch of variables that were suggested to me.:

export GPU_FORCE_64BIT_PTR=0
export GPU_MAX_HEAP_SIZE=100
export GPU_USE_SYNC_OBJECTS=1
export GPU_MAX_ALLOC_PERCENT=100
export GPU_SINGLE_ALLOC_PERCENT=100

export OMP_NUM_THREADS=1
export MPI_PER_GPU=20
export ENABLE_MPS=false
export LS_HYPERTHREAD=true

I, being a novice, do not understand how this work and I am pretty sure that the GPUs are not being utilized fully. I say this since I tried to run some basic profiling all of which indicate the same:

image
image1403×309 36.5 KB

Further, I
could not find the documentation about these.

I am using a slurm script to run my jobs. Every GPU node on the cluster has 2 Tesla V100-SXM2-16GB and 40CPU cores.

Could anyone shed some light on:

  1. What all flags and variables are there that could be set, and how to decide which are the relevant ones for my program?

  2. Where can I read more about these variables and flags?

Any insights would be much helpful.

2

You’re unlikely to get help here. This forum is for users who are making direct use of such libraries as cublas, cufft, etc. Not for QE support. There is a QE users forum here.

1 Like