Revisiting Preemptive Buffer Management for Modern Datacenter Switch Chips

Modern high-speed datacenter switch chips rely on on-chip shared packet buffers to absorb transient traffic bursts. Effective buffer management (BM) is therefore critical: BM schemes dynamically allocate shared buffer space across queues to maintain high utilization and low loss under bursty workloads. In this paper, we argue that the convergence of shallow buffers and intense traffic bursts demands a highly agile BM mechanism that can rapidly adjust allocations as congestion evolves. However, existing BMs are fundamentally constrained by their non-preemptive nature: they primarily depend on natural queue draining to reclaim buffer space, which is often too slow at modern line rates and burst timescales. This paper revisits preemptive buffer management, historically viewed as impractical in hardware, and demonstrates that it is now feasible on contemporary switch chips. We introduce Occamy, a preemptive BM design that exploits otherwise idle memory bandwidth to actively reclaim over-allocated buffer space and redistribute it to newly congested queues. Testbed experiments and large-scale simulations demonstrate that Occamy improves end-to-end performance by up to ~44%.