high performance computing on graphics processing units: hgpu.org

One of the main challenges of modern computer systems is to overcome the ever more prominent limitations of disk I/O and memory bandwidth, which today are thousands-fold slower than computational speeds. In this paper, we investigate reducing memory bandwidth and overall I/O and memory access times by using multithreaded compression and decompression of large datasets. Since the goal is to achieve a significant overall speedup of I/O, both level of compression achieved and efficiency of the compression and decompression algorithms, are of importance. Several compression methods for efficient disk access for large seismic datasets are implemented and empirically tested on on several modern CPUs and GPUs, including the Intel i7 and NVIDIA c2050 GPU. To reduce I/O time, both lossless and lossy symmetrical compression algorithms as well as hardware alternatives, are tested. Results show that I/O speedup may double by using an SSD vs. HDD disk on larger seismic datasets. Lossy methods investigated include variations of DCT-based methods in several dimensions, and combining these with lossless compression methods such as RLE (Run-Length Encoding) and Huffman encoding. Our best compression rate (0.16%) and speedups (6 for HDD and 3.2 for SSD) are achieved by using DCT in 3D and combining this with a modified RLE for lossy methods. It has an average error of 0.46% which is very acceptable for seismic applications. A simple predictive model for the execution time is also developed and shows an error of maximum 5% vs. our obtained results. It should thus be a good tool for predicting when to take advantage of multithreaded compression. This model and other techniques developed in this paper should also be applicable to several other data intensive applications.