#MoCObench-MPI-metaheuristic Optimisation algorithm for MUBQP-like problems.

##Prerequisites

• mpich
• g++
• gnuplot-x11
• gnuplot

##Compilation

mpic++ MUBQPEval-heuristic-par.cpp

##Execution

mpirun -np [number of processes] ./a.out instance.dat [solutions pool size]

The larger the number of solutions to send to the workers, the later the filtering. The smaller the number of solutions, the more important the losses in communication time.

##Interruption To interrupt the process and display efficiency informations, send a SIGUSR1 to mpirun's process:kill -s [mpirun's PID]

##Sequential version The sequential version takes the MUBQP instance as only parameter:

./MUBQPEval-heuristic-seq instance.dat

The ctrl+c interruption will display the efficiency informations.

##Deployment

• Local deployment:

./launch_instances.sh [number of instances] [instance duration (seconds)] [number of processes]
[instance file] [pool size]

• Passive Grid5000 deployment:

oarsub -l /nodes=[number of nodes],walltime=xx:xx:xx "./launch_passive_instances.sh [number of
instances] [instance duration (seconds)] [instance file] [pool size]"

Objective vectors will be stored in files named: [instance file][duration][number of processes][pool size][instance number].txt

##TODOs

• Send back solutions in blocks
• Filter whatever M (remove #defines)
• Implement compare function to refactor filter_solutions()
• Avoid copies
• Avoid returning seeds: return index of original solution, and number of flipped bit (requires a vector of sorted non-optimal-but-being-explored solutions)
• Compare with best results on MoCObench website
• Send pools of solutions (seeds) (parametrized size)
• Implement clean sequential version and compare performances
• Evaluate computation time and communication time
• Dynamic pool size: 25% ; 50% ; 75% or 100% of (number of solutions to explore) / (number of processes)
• Grid5000