This project explored how to make CXL.mem tiered memory behave more like fast DRAM by intelligently deciding when to migrate pages from slower CXL memory without relying on CPU-heavy OS techniques like page-table scanning or hinting page faults. I worked with a diverse team to develop a counting Bloom filter (CBF)–based migration policy to identify “hot” remote pages using compact, hardware-friendly approximate frequency tracking, paired with a standard LRU eviction policy to move pages back to CXL when DRAM is full.

To handle the natural saturation problem of counting Bloom filters, we implemented a dual leader/follower CBF that periodically swaps and resets filters while continuously tracking access history. The full policy was implemented in a trace-driven simulator and tested first on synthetic locality-controlled traces, then on large Intel PIN–collected server-style traces (e.g., GCC/SQLite compilation, a 1B-element sort, and multiple SPEC CPU 2017 workloads).

Results showed that Bloom filters can be a solid migration strategy and that migration policy choices have a much larger impact on bus activity/cost than swapping eviction policies (even compared to Belady’s oracle in small-trace experiments). However, on many real benchmark traces, a surprisingly simple “migrate on access” + LRU approach matched or beat the Bloom filter, suggesting that under the assumed costs aggressive cache-like migration can be competitive for workloads with weaker locality. Overall, the work demonstrates both a practical, controller-implementable hot-page detector and a clear takeaway that the right migration trigger dominates CXL efficiency, and workload locality strongly determines whether “smart” detection outperforms simpler heuristics.