Storage & ML systems

I measure how systems behave, then make the common path simpler, faster, and more efficient.

A timeline comparing exact LRU, delayed promotion, and eviction-time promotion

Latest post07.23.26

Your LRU cache is working too hard.

The archive · 2026

  1. A paper reproducibility agent: PDF in, report out We turned the MLSys'26 AE agent's core into a standalone paper reproducibility agent. Give it a paper PDF; it finds the repo, installs and runs it, and adjudicates the paper's claims.
  2. Same open weights, prompt-cache lifetimes from seconds to a day Across 23 providers serving two open-weight models, prompt-cache retention ranged from under 30 seconds to more than a day. Even for the same model, the serving stack determined the cache.
  3. Deploying an artifact-evaluation agent at MLSys'26 We deployed a tool-using agent on the MLSys'26 AE HotCRP as a real reviewer. Across its completed reviews, it usually scored within one point of its human co-reviewers.
  4. Measure first, design second AI is making solutions cheaper, but it is not making important problems easier to find. Measurement is how systems research discovers what matters.
  5. Proprietary prompt caches expire in minutes, not hours — cheap per token doesn't mean a cheap bill OpenAI, Gemini, Anthropic, and xAI all cache your prompt prefix — but the probe that caught open-weight backends holding prefixes for hours catches the big proprietary APIs letting go in 5 to 30 minutes.
  6. Authorship inflation at OSDI: counting papers two ways OSDI author lists have more than doubled since 1994. Whether the field's most prolific authors are actually publishing more depends on whether a paper counts as one, or as one divided by its author count.
  7. Make model precision elastic MorphServe temporarily converts selected model layers into KV-cache headroom when traffic spikes, then restores full precision when the pressure passes.
  8. Why FIFO is (almost) all you need for cache eviction LRU tries to rank everything. Production traces showed that caches need something simpler: reject one-hit objects quickly and delay promotion until it matters.
  9. What I look for in systems research students I am looking for curious, persistent students who want to build real systems. Here is what the work requires and how to write an email I will read carefully.

Computer Systems Researcher and Educator