An Efficient Lossless Compression Method for Internet Search Data in Hardware Accelerators

Nowadays machine learning algorithms are intensively used to improve the relevance of search engines by training on Internet search data, while their software implementations on commodity computers are not efficient (in terms of computation speed, power consumption, etc). Therefore FPGA-based accelerators have been proposed to process these large scale data. Data compression/decompression technology plays an important role for it could significantly increase those acceleratorspsila performance. Based on the statistics on datasets and the analysis of conventional data compression algorithms, we propose a bit mapping compression/decompression method to provide non-blocking streaming data to accelerators. Experiments indicate that the performance of hardware accelerator is increased up to 140% after adding the bit mapping modules. This method could also be extended to other data-intensive hardware accelerators.