Caches consume many resources in the modern datacenter, so it is important that they are used well. Unfortunately, designing an effective caching policy is hard. Applications vary tremendously in how they access data, and modern storage systems incorporate multiple technologies (e.g., DRAM, flash, disk) that introduce complex tradeoffs within and across layers of the storage hierarchy.
We propose an adaptive caching policy that learns online how to make the best use of the full storage hierarchy for different applications. The key to making this policy effective lies in our recent work on adaptive, model-driven caching policies, which gives us a method to translate simple caching statistics into meaningful caching decisions. This proposal improves on this prior work by modeling the full storage stack and employing machine learning to improve caching decisions.
Nathan Beckmann, Phillip Gibbons, Bernhard Haeupler, Charles McGuffey. APoCS 2020. (Best Paper.)
Nathan Beckmann, Phillip Gibbons, Bernhard Haeupler, Charles McGuffey. SPAA 2019. (Brief Announcement.)