← All publications
MICRO2026To appear

Rethinking Compression for CXL Memory Expanders at Hyperscale

Haneul Park, Grant Ayers, Nam Sung Kim, Philip Levis, Brian Morris

To appear in the 59th IEEE/ACM International Symposium on Microarchitecture (MICRO), 2026

Host CPU 64 B read CXL CXL Memory Expander decompression engine on the critical path of every read knob 1 algorithm knob 2 block size read amp. DDR DRAM compressed blocks stronger compression → fewer DRAM bytes per access → less bandwidth and queueing

Every 64-byte read decompresses a whole block — so the algorithm and block size decide both how much DRAM traffic each access generates and how long decompression takes. (Illustrative diagram; replace with the paper's figure.)

TL;DR

Commercial CXL memory expanders default to fast, weak compression to keep decompression off the critical path. At hyperscale that is the wrong call: memory bandwidth and queueing dominate end-to-end latency, so a slower but stronger algorithm makes the expander both bigger and faster.

Why it matters

DRAM dominates the cost of hyperscale datacenter servers, and AI-driven memory supply shortages keep pushing costs higher. The OCP CXL Tiered Memory Expander cuts server cost by reusing legacy DDR modules, and inline compression stretches effective capacity further — but its compression design points were set by conventional wisdom, not measurement.

What's in the paper

This work was done during my student researcher role in Google's Systems Research Group. An early letter version appeared in IEEE Computer Architecture Letters.

BibTeX

@inproceedings{park2026rethinking,
  author    = {Haneul Park and Grant Ayers and Nam Sung Kim and Philip Levis
               and Brian Morris},
  title     = {Rethinking Compression for {CXL} Memory Expanders at Hyperscale},
  booktitle = {IEEE/ACM International Symposium on Microarchitecture (MICRO)},
  year      = {2026}
}