Blog

Notes on performance modeling, information freshness, and large language model inference

An LLM server in three numbers

A note on recent work with Asser Tantawi. A queueing model of LLM inference with three hardware parameters, accurate enough to drive an autoscaler and simple enough to tune itself.

14 min read · July 18, 2026

2026 · llm-serving queueing performance-modeling autoscaling · research
When a pipeline should stop working

Part three of a second series on my PhD work. In multi-step update processing, busy servers can waste power on results that make no one better informed.

17 min read · July 18, 2026

2026 · age-of-information queueing performance-modeling energy · research
The best time to read memory

Part two of a second series on my PhD work. The optimal policy for sampling shared memory turns out to be a threshold with a closed form.

11 min read · July 18, 2026

2026 · age-of-information memory-access performance-modeling · research
Sometimes fresher means waiting

Part one of a second series on my PhD work. Why a decision process that reads shared data and computes on it should sometimes sit idle before reading again.

11 min read · July 18, 2026

2026 · age-of-information update-processing queueing · research
Freshness has a memory bill

Part three of a series on my PhD work. How much memory Read-Copy-Update spends to keep information fresh, and why the bill stays bounded.

9 min read · July 17, 2026

2026 · age-of-information rcu shared-memory · research