A QAP Solver with CUDA GPU Computing Architecture A Two Page Description of the Application Submitted for GECCO 2009 Competition : GPUs for Genetic and Evolutionary Computation

This application solves the quadratic assignment problem (QAP) [1]. In QAP, we are given l locations and l facilities and the task is to assign the facilities to the locations to minimize the cost. We chose QAP for the following reasons: First, problem sizes of QAPs in real life problems are relatively small compared with other problems in permutation domains such as the traveling salesman problem (TSP) and the scheduling problem. This enables us to use the shared memory of a GPU e ectively. Second, QAP is one of the most di cult problems among problems in permutation domains. Thus, QAP is a good test bed to evaluate an optimization algorithm. As the parallel method, a multiple-population, coarsegrained GA model was adopted. Each subpopulation is evolved by a multiprocessor in a CUDA GPU. At predetermined intervals of generations all individuals in subpopulations are shu ed via the VRAM of the GPU. Applying local search in solving QAP is very common in evolutionary algorithms [2, 3, 4]. Popular local searches used in solving QAP are the taboo search (TS) and 2-OPT heuristic. However, to apply these local searches e ciently, we need a large memory for each parallel thread. For example, the size needed for taboo search [5] is 32l + l bytes. To take this size space for each thread in the shared memory with current GPU is almost impossible. Thus, we will not apply any local search for GPU. We applied local searches (2 OPT Best Move and 2 OPT First Move) for CPU. The instances on which this algorithm was tested were taken from the QAPLIB benchmark library at [6]. Our results were promising, showing the performance of GPU programs without local searches is comparative to or even better than the performance of CPU programs with local searches on the Intel R ⃝ Core i7 965 processor, which is one of the fastest processors available currently.