Breaking the Ice: Analyzing Cold Start Latency in vLLM
MLSys 2026Huzaifa Shaaban Kabakibo, Animesh Trivedi, Lin Wang
PhD Student · Paderborn University
LLM Systems Research·Computer Networks Group
I optimize large language model serving systems: the scheduling, autoscaling, and GPU sharing that maximize utilization and make inference cheaper to run at scale.
I'm a PhD student in the Computer Networks Group at Paderborn University, working on how to run large language model inference more efficiently at the systems level. My current focus is optimizing the serving stack itself: scheduling requests, autoscaling replicas, and multiplexing models across GPUs to squeeze more useful work out of scarce accelerator time. An earlier thread in the same direction looked at cold‑start latency in vLLM: predicting how long an engine takes to spin up so a scheduler can plan around it.
Before starting the PhD I built MoRAGBench, the first benchmarking framework for retrieval‑augmented generation on mobile devices, and worked on running LLM inference efficiently on constrained edge hardware. Across all of it, I'm drawn to the same set of problems: scalable AI systems, systems optimization, and hardware–software co‑design.
Huzaifa Shaaban Kabakibo, Animesh Trivedi, Lin Wang
Huzaifa Shaaban Kabakibo, Eric Schniedermeyer, Artem Burchanow, Lin Wang
Benchmarking framework for RAG pipelines on mobile devices
Paderborn University
Paderborn University
Lebanon
Paderborn University · Computer Networks Group — LLM serving systems & inference optimization
Paderborn University · Embedded Systems focus · Graduated with distinction (sehr gut, 1.1)
Lebanese International University · Distinguished degree, GPA 3.97 / 4.0
Tools & Systems
Languages