Quiver: Dynamic Quantile-Assisted Scheduling to Reduce Rank Inversion

Abstract

Programmable packet scheduling is essential for supporting diverse Quality-of-Service demands in modern networks. The Push-In First-Out (PIFO) queue offers an abstraction for programmable packet scheduling. However, achieving a line-rate and scalable implementation remains an open challenge. Consequently, prior research has explored approximating PIFO semantics using commodity Strict Priority (SP) queues. Existing approaches predominantly employ static or heuristic-based rank-to-queue mappings, which often lead to scheduling inaccuracies manifested as rank inversions.In this paper, we present Quiver, a dynamic quantile-assisted scheduler that continuously adapts to traffic by sampling packet ranks and computing statistical bounds for strict priority queues, minimising rank inversions at line rate.We evaluate Quiver using the NetBench simulator across diverse traffic distributions. Results show that its quantile-based design closely matches the performance of the optimal static Gradient-based algorithm while significantly reducing rank inversions. Furthermore, Quiver drops fewer packets of high-priority flows compared to existing designs and achieves Flow Completion Times (FCTs) comparable to an ideal PIFO. © 2026 IFIP.