The Harvard Systems Group at the John A. Paulson School of Engineering and Applied Sciences (SEAS) operates at the vital intersection of systems, theory, and hardware. We design and build the architectures, networks, and databases that form the backbone of modern computing.
From scaling global cloud infrastructure to optimizing the software-hardware interface, we don’t just study systems—we reinvent the layers that power the future.
Humans naturally reason about relations such as cause and effect, and this reasoning is both bidirectional and compositional. Matthew will present energy-based relational programming, a purely neural framework that represents each relation with a smoothed energy function and composes them zero-shot at query time to solve reasoning problems.
Large language models increasingly rely on multiple GPUs, but communication overhead and synchronization limit their efficiency. Yang will present UCCL, a family of systems that improves LLM performance through efficient parallelism and extensible networking, reducing inference latency and improving communication throughput.
Distributed AI workloads use the network in unique ways, so traditional networking solutions are poorly suited to interconnecting the massive number of processors in AI training and inference clusters. Mario will analyze these networking requirements and describe the architectures and standard protocol stacks built for scale-out and scale-up AI clusters, concluding with how they come together in a commercial solution.
Smart cache eviction algorithms can adapt to workloads, but they often pay for it with complexity, instability, or overhead. Learning-Augmented Heuristics takes a different approach: keep the fast heuristic on the data path, and use learning to configure it at a slower timescale.
Bursty arrivals usually cause a drop in performance and are seen as headaches in production systems. In this blog, we investigate a phenomenon where burstiness actually improves performance, uncovering what conditions produce this effect.
Cache eviction is usually presented as a ranking problem. LRU ranks objects by
recency. LFU ranks them by frequency. More advanced algorithms combine several
signals, adapt their parameters, and maintain increasingly sophisticated data
structures to decide which object is least valuable.
Most LLM servers treat a deployed model as fixed: its weights have one
precision, its KV cache has one memory budget, and both remain unchanged until
the process restarts.
