Benchmark: Bio-stochastic Load-imbalance for Multi-GPUs

This repository contains a benchmark and benchmarking script for multiple GPU task-to-gpu (task2gpu) scheduling strategies applicable for stochastic applications with load imbalance in computational biology.

It includes 3 different task2gpu scheduling approaches:

1. dynamic_rand - An asynchronous method, it choses GPU randomly.
2. dynamic_occ - A synchronous method works based on the occupancy of the GPU.
3. dynamic_ad - An asynchronous method works based on the load of the GPU.

Build

Can use the sample script compile_summit.sh if you are using OLCF Summit, using llvm clang or IBM xl compiler.

# using llvm
module use /sw/summit/modulefiles/ums/stf010/Core
module load cuda llvm/12.0.0-20210125
clang++ -Ofast -std=c++17 -fopenmp  -fopenmp-targets=nvptx64-nvida-cuda --cuda-path=${OLCF_CUDA_ROOT} task2gpu_bench.cpp -o t2g 

# using xl
module load xl/16.1.1-10
xlc -std=c++17 -qsmp=omp -qoffload task2gpu_bench.cpp -o t2g

Run

Can use the sample script run_summit.sh in the ineractive mode if you are using OLCF Summit.

jsrun --smpiargs="-disable_gpu_hooks" -n 1 -r 1 -a 1 -g 6 -c 42 -d packed -b rs -l CPU-CPU ./t2g 256 512 1 10  

Parameter list

upper limit of task size, number of tasks, grain/chunk size, and upper limit of the convergent rate