Compress objects, not cache lines: an object-based compressed memory hierarchy

Finally, if we also use cross-object compression in which a base object is stored for each object type, and only the delta from the base object is stored for every instance, then we can get the compression ratio up as high as 95% (d)! Zippads addresses this by using different compression algorithms: COCO for objects and a conventional BDI+FPC algorithm for arrays. In the figure below, CMH is a state-of-art compressed memory hierarchy, hotpads is plain hotpads (no compression) and a three-level cache hierarchy, Zippads-BF is Zippads without COCO, and Zippads is the full Zippads with dynamic selection of compression algorithm including COCO.

Source: blog.acolyer.org