Seminars
Rethinking Bursty Workloads and KV Cache Hierarchies for Efficient LLM Serving
Akira will present two recent projects on serving bursty, prefix-heavy LLM inference workloads: one showing how bursty arrivals can improve a cluster’s time-per-output-token and what that implies for request routing, and another on tuning the private vs. shared split of a distributed KV cache to optimize time-to-first-token—together demonstrating that intuitive systems choices can leave performance gains on the table.
AI and Dependability in Computing Systems: Friend or Foe?
Saurabh will examine both sides of AI’s effect on the reliability and security of computing systems—a foe when it complicates reasoning about probabilistic programs and enables phishing at scale, a friend when it powers causal root-cause diagnosis and automatic vulnerability patching—and highlight recent work from the NSF Center Chorus on streaming ML with probabilistic accuracy guarantees and secure decentralized ML.
Network-aware co-design for distributed machine learning systems
As ML training scales across racks, data-center buildings, regions, and even federated clients, periodic synchronization leaves compute idle and prolongs iterations. Minghao will argue for network-aware co-design—jointly optimizing algorithms, parallelism and pipeline strategies, and heterogeneity handling—and present THC, ScaleAcross Explorer, and FIELDING as works that cut synchronization costs across increasingly large deployments.
Simple Techniques for Loop-friendly Eviction and Belady Anomaly Reduction
Many workloads, such as repeated scans and LSM-tree compaction, access groups of objects in similar orders, forming access loops that several cache eviction algorithms handle well—though it remains unclear what makes them effective. Yunjia will distill the structural features behind this efficiency into portable eviction gadgets that can be plugged into existing policies, and show that these features also improve cache stability by reducing miss-ratio anomalies and cliffs.
Diagnosing Performance Issues by Uncovering Application Resource Bottlenecks
Performance issues in large software systems are often driven by application-defined resources like buffer pools and caches, which remain hidden from system-level metrics and are notoriously difficult to diagnose with standard profilers. Youliang will introduce gigiprofiler, a tool that combines LLM-based semantic inference with static analysis to uncover these bottlenecks, and share practical insights on leveraging large language models for program analysis.
A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing
Designing effective modern LLM serving systems requires an understanding of realistic workloads, but capturing the complexity of today’s diverse applications is difficult using only short traces or synthetic datasets. William will share insights from a comprehensive one-year production trace of billions of LLM requests, exploring how these workloads evolve and detailing key systems implications for prefix caching and load balancing.
Sketches and Their Applications for Synchronization, Blockchain Networks and Learning
Hash-based data structures such as Bloom filters are widely used in network systems for a wide range of tasks. In this talk, Ori will overview several recent designs that expand and enhance their utility across multiple domains such as data synchronization among peers, blockchain networks and machine learning pipelines. The talk is based on recent papers from Sigmetrics, TNSM, AFT and APNET.
Systems Seminar: Round Table Discussion
A new addition to our Systems Seminar: an open round table discussion where people are encouraged to share recent papers, tech news, new tools, or open questions in a relaxed, conversational setting.
Building Scalable Distributed Databases in the Age of Geo-Replication
Modern distributed applications depend heavily on geo-replication for fault tolerance, but high network latency forces these databases to make difficult tradeoffs between the high performance of weak consistency and the data safety of strong consistency. Yunhao will explore solutions to these challenges by detailing advancements in Conflict-free Replicated Datatypes (CRDTs), including the Janus implementation, and introducing Minerva, a scalable transaction protocol designed to maintain high throughput across wide-area networks.
Systems at the Crossroad of Agents & Infrastructure (MLSys ’26 Digest Talk)
As Large Language Models transition into autonomous agentic systems, traditional serving frameworks are facing unprecedented performance bottlenecks. Yiyu will deliver a structured digest of cutting-edge research from MLSys 2026 covering agentic AI, LLM systems, and compilers.