Undergraduate researcherYonsei University
Yunmin Cha
I build cross-layer systems for predictable and efficient LLM inference.
My systems work spans the inference stack: how a GPU is shared among workloads, which code actually determines performance, and how data reaches GPU memory. A second line of research asks when AI assistance improves business decisions.
Layer 1 · Requests & SLOs
Which request should run next?
Requests arrive with latency targets, their service-level objectives (SLOs), and some wait close to missing them. Choosing what runs next is a decision under uncertainty: no one knows how long each response will be.
Layer 2 · Serving runtime
How should one GPU be divided?
The serving engine batches requests and keeps their state in GPU memory, allocating and freeing pages as they come and go. I work on the scheduling layer beneath it, which divides one GPU among several workloads.
Layer 3 · GPU execution
Which code actually sets the speed?
When a request and its data arrive, kernels run across the GPU’s lanes. A profiler shows where time goes, not which change would make the application faster; causal analysis estimates that by experimenting on the running program.
Layer 4 · Memory & storage
Who should move the data?
As models outgrow GPU memory, data must stream in from SSDs. Whether the CPU or the GPU controls that movement shapes how fast it arrives; I compare these designs on the same machine.
