Towards Compression At All Levels In The Memory Hierarchy

Hardware compression techniques are typically simplifications of software compression methods. They must, however, comply with area, power and latency constraints. This study unveils the challenges of adopting compression in memory design. The goal of this analysis is not to summarize proposals, but to put in evidence the solutions they employ to handle those challenges. An in-depth description of the main characteristics of multiple methods is provided, as well as criteria that can be used as a basis for the assessment of such schemes. Typically, these schemes are not very efficient, and those that do compress well decompress slowly. This work explores their granularity to redefine their perspectives and improve their efficiency, through a concept called Region-Chunk compression. Its goal is to achieve low (good) compression ratio and fast decompression latency. The key observation is that by further sub-dividing the chunks of data being compressed one can reduce data duplication. This concept can be applied to several previously proposed compressors, resulting in a reduction of their average compressed size. In particular, a single-cycle-decompression compressor is boosted to reach a compressibility level competitive to state-of-the-art proposals. Finally, to increase the probability of successfully co-allocating compressed lines, Pairwise Space Sharing (PSS) is proposed. PSS can be applied orthogonally to compaction methods at no extra latency penalty, and with a cost-effective metadata overhead. The proposed system (Region-Chunk+PSS) further enhances the normalized average cache capacity by 2.7% (geometric mean), while featuring short decompression latency.