1
RESEARCH POSTER PRESENTATION DESIGN © 2011 www.PosterPresentation s.com • Genome size is a parameter that determines the length of the genetic information carried by individuals. • Increasing the genome size causes the algorithm to execute longer for both CPU and GPU algorithms since genome size is directly proportional to the size of data processed. In this paper, a parallel memetic algorithm implementation for CUDA platform is described. The conventional genetic operators are adapted to the GPU considering the GPU architecture. In this population based optimization technique, there are one more islands and each island consists of constant number of individuals. Each CUDA thread is responsible for evolution of one individual, and islands are mapped as CUDA blocks to benefit from the shared memory. The results show up to 38x speedup compared to the CPU implementation. ABSTRACT Populations reside in thread blocks, so the shared memory is effectively used for genetic operators including crossover, mutation and local search. In each thread block, each thread is responsible for an individual. At the beginning of the kernel code, each thread reads its genome data from global memory and puts them into shared memory. After that point all of the calculations done by threads are made on shared memory except the migration operator. Since migration operator requires communication between distinct islands (thread blocks), global memory has to be employed for this operation. GENERAL DESIGN KERNEL PSEUDO CODE EXPERIMENTS TESTING CONFIGURATION The performance of our implementation has been evaluated on a PC having Intel Core i7 870 @ 2.93 GHz processor and NVDIA GeForce GTX 460 (336 cores) GPU. All of the experiments have been repeated 20 times using the same configuration.. The parallel memetic algorithm configuration: • population size: 16 to 512 • number of islands: 16 to 1024 • genome size: 4 • tournament selection with winning probability of 0.8 • mutation probability: 0.05 • local search iteration number: 10 • local search step size: 0.001 • migration interval: 10 CONCLUSION • While it is dependent on the fitness functions, GPU implementation shows better performance for all types of fitness functions. • The best speedup is achieved with Griewangk function where 38x speedup is observed. • The future work will be focusing on more complex optimization problems providing solutions to real world problems. • Moreover, different genetic operators (local search, mutation and crossover) are planned to be implemented on CUDA to achieve better optimization results. Parallel Memetic Algorithm Implementation on CUDA Umut Cinar, Alptekin Temizel Graduate School of Informatics, Middle East Technical University, Ankara, Turkey Function Formula Global Minimum Properties Rosenbrock = [100 +1 − 2 2 + ( − 1) 2 ] −1 =1 ∗ = 1,1, … , 1 ; ∗ =0 Not seperable Rastrigin = 10 + [ 2 − 10cos(2 )] =1 ∗ = 0,0, … , 0 ; ∗ =0 Seperable and Multimodal Griewangk =1+ 2 4000 =1 − cos( ) =1 ∗ = 0,0, … , 0 ; ∗ =0 Not seperable and Multimodal Schwefel ℎ = ( =1 ) 2 =1 ∗ = 0,0, … , 0 ; ℎ ∗ =0 Not seperable Initialize shared memory buffers WHILE termination criterion is not met; DO evaluate individual begin local search tournament selection IF thread_id < warp_size offsprings = crossover(parents) offsprings = mutation(offsprings) replace offsprings with worst_individuals ELSE IF thread_id < warp_size*2 AND (MIGRATION_INTERVAL is true) migrate(best_individuals) END IF END WHILE • Threads within a block continues in parallel until the variation and migration operators take place. • At this point, two warp of threads executes crossover/mutation and migration operations. First warp is assigned to crossover/mutation while the second warp is assigned to migration (when it is migration interval). • Hence, this part of the implementation is completely unrolled by using a warp for each operation. This approach decreases the cost of warp divergence significantly. • The experiments show that the results are heavily dependent on the fitness function type. In a memetic algorithm, the most computationally complex routine is the fitness function, and this is reflected in the dependency of performance on fitness function type. • The maximum speedups achieved with Griewangk function as x38. 16 64 256 0 2 4 6 8 10 12 14 Speedup Population Size Number of Islands Rosenbrock function on GPU (max speedup is x13) 12-14 10-12 8-10 6-8 4-6 2-4 0-2 16 64 256 0 5 10 15 20 25 Speedup Population Size Number of Islands Rastrigin function on GPU (max speedup is x25) 20-25 15-20 10-15 5-10 0-5 16 64 256 0 2 4 6 8 10 Speedup Population Size Number of Islands Schwefel function on GPU (max speedup is x9) 8-10 6-8 4-6 2-4 0-2 16 64 256 0 10 20 30 40 Speedup Population Size Number of Islands Grienwangk function on GPU (max speedup is x38) 30-40 20-30 10-20 0-10 Speedup has been investigated using several benchmarking functions including Rosenbrock, Rastrigin, Schwefel and Griewangk and various parameters. INFLUENCE OF GENOME SIZE ON EXECUTION TIME 53 97 141 185 229 1989 3408 5027 6626 8536 1 10 100 1000 10000 2 4 6 8 10 Execution Time (ms) Genome Size Genome Size vs Execution Time CPU GPU REFERENCES [1] Pablo Moscato, Carlos Cotta, “A Gentle Introduction to Memetic Algorithms”, Handbook of Metaheuristics, pp. 105-144, 2003. [2] Wojciech Bozejko, Mieczyslaw Wodecki, “The Methodology of Parallel Memetic Algorithms Designing”, in Proc. ICAART 2011.
![Comparative Genome and Proteome Analysis of Anopheles ...cegg.unige.ch/system/files/149.pdf · genome size [the Anopheles genome is twice the size of that of Drosophila (3–5)].](https://static.documents.pub/doc/80x56/5f0a4ff27e708231d42b09b1/comparative-genome-and-proteome-analysis-of-anopheles-ceggunigechsystemfiles149pdf.jpg)
