Section: New Results

Multi-core GPU-based parallel optimization

We have mainly investigated the design and implementation on multi-core GPU-based platforms of metaheuristics and tree-based exact optimization methods focusing on Branch and bound (B&B) algorithms (Ph.D thesis of I. Chakroun).

• GPU-based Metaheuristics

  Participants: N. Melab, T-V. Luong, K. Boufaras and N. Melab

  We came out with a pioneering work on single-solution methods. The hierarchy of parallel models has been rethought on GPU dealing with CPU-GPU data transfer optimization, thread control and automatic mapping of candidate solutions to threads. The implementation of the proposed approaches is provided through ParadisEO-GPU in [62] (nominated for Best Paper Award). High speedups ups have been achieved for some problems.

• Multi-core GPU-based B&B algorithms

  For exact optimization, we have revisited the design and implementation of highly irregular B&B algorithms on GPU dealing with hierarchical device memory optimization, on GPU combined with multi-core [45] dealing with CPU-GPU data transfer optimization and work partitioning, and on GPU-enhanced computational grids. Accelerations up to $\times 217$ are achieved on Tesla Nvidia C2050 on large Flow-Shop problems.