Implementation of Genetic Algorithms in FPGA-based Reconfigurable Computing Systems

Genetic Algorithms (GAs) are used to solve many optimization problems in science and engineering. GA is a heuristics approach which relies largely on random numbers to determine the approximate solution of an optimization problem. We use the Mersenne Twister Algorithm (MTA) to generate a non-overlapping sequence of random numbers with a period of 219937-1. The random numbers are generated from a state vector that consists of 624 elements. Our work on state vector generation and the GA implementation targets the solution of a flow-line scheduling problem where the flow-lines have jobs to process and the goal is to find a suitable completion time for all jobs using a GA. The state vector generation algorithm (MTA) performs poorly in traditional von Neumann architectures due to its poor temporal and spatial locality. Therefore its performance is limited by the speed at which we can access memory. With an approximate increase of processor performance by 60% per year and a drop of memory latency only 7% per year, a new approach is needed for performance improvement. On the other hand, the GA implementation in a general-purpose microprocessor, though performs reasonably well, has scope for performance gain in a parallel implementation. The parallel implementation of the GA can work as a kernel for applications that uses a GA to reach a solution. Our approach is to implement the state vector generation process and the GA in an FPGA-based Reconfigurable Computing (RC) system with the goal of improving the overall performance. Application design for FPGA-based RC systems is not trivial and the performance improvement is not guaranteed. Designing for RC systems requires algorithmic parallelism in order to exploit the inherent parallelism of the FPGA. We are using a high-level language that provides a level of abstraction from the lower-level hardware in the RC system making it difficult to fully exploit some of the architectural benefits of the FPGA. Considering these factors, we improve the state vector generation process algorithmically. Our implementation generates state vectors 5X faster than the previous implementation in an Intel Xeon microprocessor of 2GHz. The modified algorithm is also implemented in a Xilinx Virtex-4 FPGA that results in a 2.4X speedup. Improvement in this preprocessing step accelerates GA application performance as random numbers are generated from these state vectors for the genetic operators. We simulate the basic operations of a GA in an FPGA to study its behavior in a parallel environment and analyze the results. The initial FPGA implementation of the GA runs about 7X slower than its microprocessor counterpart. The reasons are explained along with suggestions for improvement and future work