Reproduce “DPC: A Distributed Page Cache over CXL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shai Bergman, Zhe Yang, Julien Eudine, Giorgio Negro, Onur Mutlu, Arash Tavakkol, Ji Zhang
DPC is an OS-level distributed page cache that leverages CXL 3.0 to unify cluster memory into a single, coherent cache, addressing inefficiencies of per-node caches. It demonstrated up to 16.2x speedup for Webserver and a 5.6x geometric mean speedup across various data-sharing workloads by reducing redundant data and facilitating low-latency remote cache hits.
Key insights
- CXL 3.0's hardware-managed coherence and remote memory access capabilities can be leveraged at the OS level to transform independent node caches into a unified, cluster-wide page cache.
- Enforcing a single-copy invariant for cached data pages across a cluster significantly improves aggregate DRAM utilization by preventing redundant data replication.
- A hybrid approach combining CXL hardware for low-latency remote memory access with a software-defined page-granular directory protocol enables scalable coherence for distributed caching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.