A P4/BMv2 Proof-of-Concept Reimplementation and Assessment of PIAS Flow Scheduling

Abstract

Data center applications often generate highly skewed flow-size distributions, where short flows are latencysensitive and can be adversely affected by queueing behind long flows. PIAS [1] addresses this problem through an informationagnostic scheduling approach that approximates shortest-job-first behavior without requiring prior knowledge of individual flow sizes. Instead, PIAS tags packets according to the amount of data already sent by each flow and relies on strict-priority queues in commodity switches. This paper presents a proof-of-concept reimplementation of PIAS using P4, BMv2, and Mininet. The artifact implements DSCP-based priority tagging, BMv2 strictpriority queueing, workload-derived demotion thresholds, and ECN marking. The evaluation considers a single-ToR topology with 16 hosts, a web-search workload, multiple offered loads, and seven random seeds per configuration. Results show that the reimplemented PIAS mechanism substantially reduces flow completion time relative to FIFO under matched thresholds. At 0.8 offered load, mean FCT for small flows decreases from 416.81 ms with FIFO to 17.99 ms with PIAS and ECN enabled. Queuecount experiments further show that increasing the number of priority queues improves FCT, especially at high load. The implementation also validates ECN CE marking in the P4 data plane, although the observed transport-level FCT benefit of ECN is not uniform in the Mininet/Linux-TCP environment. These results support the feasibility of using P4/BMv2 as an experimental platform for studying PIAS-like flow scheduling, while also identifying limitations that must be considered when interpreting software-switch emulation results.