BenchParallelMeta

This repository contains the source code for the comparative study presented in the paper

Jan Gmys, Tiago Carneiro, Nouredine Melab, El-Ghazali Talbi, Daniel Tuyttens, A comparative study of high-productivity high-performance programming languages for parallel metaheuristics, Swarm and Evolutionary Computation

=================

Content

Parallel implementations in C++/OpenMP, Julia, Python-Numba and Chapel:

1. For the 3-dimensional Quadratic Assignment Problem (Q3AP)

   • Q3AP-ILS: Iterated Local Search (ILS) metaheuristic using multi-threaded parallel neighborhood evaluations.
   • Q3AP-GA: Genetic Algorithm, hybridized with Parallel Local Search.

2. ParallelBatchEval: A micro-benchmark evaluating the performance of parallel batch evaluation of solutions for different batchsizes and different cost functions:

   • Permutation Flowshop Scheduling Problem (PFSP)
   • Quadratic Assignment Problem (QAP)
   • Traveling Salesman Problem (TSP)
   • Q3AP

=================

How to run

ILS

Python

• no numba, instance nug12-d, 10 ILS iterations
  • python3 ./ilsQ3AP_numpy.py nug12 10
• no numba + multiprocessing
  • python3 ./ilsQ3AP_numpy_par.py nug12 10
• with numba
  • sequential: python3 ./ilsQ3AP_numba_seq.py nug12 100
  • parallel: python3 ./ilsQ3AP_numba_seq.py nug12 100

To set the number of threads in the parallel numba version

• ex.: export NUMBA_NUM_THREADS=4
• OpenMP environment variables should be unset as they may interfere

Julia

• sequential (nug12-d, 100 iterations)

  • julia ./ils_q3ap_seq.jl nug12 100

• parallel (8 threads)

  • export JULIA_NUM_THREADS=8
  • julia ./ils_q3ap_par.jl nug12 100

C

• compile:
  • make
• run with 8 threads (nug12-d, 100 iterations)
  • OMP_NUM_THREADS=8 ./ils nug12 100

---

ParallelBatchEval

Python

3 arguments required: problem | instance| batchsize

Examples:

• python3 ./bench.py TSP berlin52 1000
• python3 ./bench.py FSP ta020 10000
• python3 ./bench.py QAP nug15 100
• python3 ./bench.py Q3AP nug12 1000

Julia

3 arguments required: problem | instance| batchsize

Examples:

• julia ./bench.jl TSP berlin52 1000

C

• make
• ./bench -z p=fsp,i=ta120 -b 1000
• ./bench --help gives some help