Title: Simplifying Prioritization and Scheduling with P²CS
Authors: Ali Munir (Huawei Technologies Co., Ltd.); Xiaolin Pang (Huawei Technologies Co., Ltd.); Junyi Zhang (Huawei Technologies Co., Ltd.)
Introduction
Modern datacenter networks host a diverse mix of applications and services with varying quality‑of‑service (QoS) requirements, yet commodity switches are limited to only eight physical priority queues. This hardware constraint severely restricts both traffic isolation across different workloads and fine‑grained prioritization within the same traffic class—capabilities that are essential for latency‑sensitive services, distributed machine learning (DML) training, and throughput‑intensive backend operations. Existing solutions fall into two categories: end‑host‑based approaches that require global priority information, which is difficult to obtain and maintain at scale, and in‑network mechanisms that demand complex sorting and scheduling logic under stringent hardware constraints. Both struggle to bridge the gap between theoretical scheduling performance and practical deployability, leaving network operators unable to effectively prioritize traffic without costly hardware upgrades or significant complexity overhead.
Key Idea and Contribution
The authors introduce P²CS (Priority‑based Probabilistic Congestion Signaling), a lightweight and practical approach that achieves fine‑grained traffic prioritization using only a single FIFO queue. P²CS operates through three coordinated components: (i) priority‑aware congestion signaling, which modulates ECN/CNP marking probabilities based on packet priority, causing low‑priority flows to reduce their sending rates more aggressively; (ii) priority‑aware packet dropping, which probabilistically discards low‑priority packets when switch buffers approach capacity; and (iii) optional switch‑side priority arbitration, which maintains a simple per‑port register to track the highest active priority locally, approximating strict priority without complex sorting hardware.
The key innovation is that P²CS shifts prioritization logic from switch packet ordering to end‑host congestion control, leveraging widely supported congestion signals to enforce priority distinctions. This design achieves three critical objectives: multi‑level prioritization via a single queue without priority inversion or out‑of‑order delivery; work‑conserving behavior that prevents starvation while maintaining full bandwidth utilization; and immediate deployability on commodity switches with only modest modifications—approximately 40 lines of P4 code for ECN integration. P²CS supports fair, weighted, and strict priority regimes through configurable priority tagging and switch‑side probability distributions, and can be combined with physical priority queues to enable hybrid scheduling strategies for scenarios such as multi‑tenant ML workloads and mixed ECMP/packet‑spraying traffic.
Evaluation
The authors evaluate P²CS through both hardware testbed experiments using Tofino P4 switches and large‑scale ns‑3 simulations across representative datacenter workloads. In testbed experiments, varying ECN marking probabilities proportionally controls flow throughput, effectively emulating Weighted Fair Queuing and Strict Priority Queuing with a single FIFO queue. For distributed ML training with background traffic, P²CS improves iteration time from 1.53 ms under fair sharing to 0.84 ms under strict priority—approaching the ideal completion time without background traffic. In large‑scale simulations of websearch and datamining workloads, P²CS using a single physical queue outperforms end‑host‑based mechanisms like L²DCT and achieves performance within 4–8 % of PIAS, which requires eight physical queues. When combined with physical priority queues, P²CS+ approaches or exceeds the performance of in‑network mechanisms such as pFabric while reducing packet drops by orders of magnitude. For multi‑tenant DML workloads, P²CS simultaneously achieves task interleaving and tenant isolation with only two physical queues, closely matching the ideal performance that would require four precious priority queues. For mixed ECMP and packet‑spraying traffic, P²CS improves ECMP flow throughput by 17 % without degrading spray traffic performance—a result that cannot be achieved through hardware‑based priority queues alone. This result is significant because it demonstrates that a lightweight, probabilistically driven approach can match or surpass complex in‑network scheduling mechanisms, enabling network operators to achieve fine‑grained prioritization without costly hardware upgrades or complex configuration.
Personal Thoughts
This paper presents a refreshingly pragmatic solution to a long‑standing problem in datacenter networking. What I find most compelling is its elegant insight: prioritization need not be enforced through complex dequeue scheduling but can instead be achieved through probabilistic congestion signaling—a fundamentally simpler mechanism that leverages existing ECN infrastructure. The authors convincingly show that when paired with fast‑converging congestion control protocols, admission control and rate regulation can be as effective as service ordering, especially given the shallow queue depths maintained by modern transports. The practical deployability of P²CS, requiring only minimal P4 modifications and no changes to existing ECN configurations, is a major strength that bridges the gap between theory and practice. I also appreciate the authors’ honesty in discussing limitations, including the inability to guarantee strict priority for very small flows completing within a single RTT and the dependence on congestion control algorithms that react proportionally to per‑packet feedback.