Accelerating disk access using compression for large seismic datasets on modern GPU and CPU
FREE-DOWNLOAD AA Aqrawi 2010
One of the major challenges of modern architectures is to overcome the limitations of disk and memory bandwidth, which per today are much slower than computation speeds.In this paper, several compression methods for efﬁcient disk access on both the CPU and GPU are described and empirically tested. To reduce I/O time we have tested both lossless and lossy compression algorithms and hardware alternatives. Our results show that an I/O speedup of 2 is achieved by using an SSD vs. HDD disk on seismic data. The use of compression for I/O gave a speedup of 1.08 and 1.2 with lossless compression methods of RLE and Huffman, respectaivly, and up to 6 for lossy methods with an average error of 0.46%. Lossy methods include performing variations of DCT, in several dimensions, and combining these with lossless compression methods such as RLE and Huffman. The speedup was achieved by enabling collaboration between the CPU and GPU.