[Paper Review] Datacenter Ethernet and RDMA: Issues at Hyperscale

Torsten Hoefler, Duncan Roweth+10 more|arXiv (Cornell University)|Feb 7, 2023

Software-Defined Networks and 5G4 citations

TL;DR

This paper argues that RoCE, despite its widespread use, suffers from fundamental flaws in lossless networking, congestion control, and scalability at hyperscale. It proposes a next-generation Ethernet-based high-performance networking stack that supports lossy and lossless modes, fine-grained retransmission (e.g., selective ACKs), and integrated congestion control to replace outdated RoCE and TCP/IP models for AI, HPC, and storage workloads.

ABSTRACT

We observe that emerging artificial intelligence, high-performance computing, and storage workloads pose new challenges for large-scale datacenter networking. RDMA over Converged Ethernet (RoCE) was an attempt to adopt modern Remote Direct Memory Access (RDMA) features into existing Ethernet installations. Now, a decade later, we revisit RoCE's design points and conclude that several of its shortcomings must be addressed to fulfill the demands of hyperscale datacenters. We predict that both the datacenter and high-performance computing markets will converge and adopt modernized Ethernet-based high-performance networking solutions that will replace TCP and RoCE within a decade.

Motivation & Objective

• To identify critical shortcomings of RoCE in modern hyperscale datacenters, particularly in congestion control, lossless operation, and scalability.
• To analyze how legacy RDMA semantics—such as in-order delivery and go-back-n retransmission—hinder efficient load balancing and network utilization.
• To argue for a fundamental redesign of datacenter Ethernet to support both lossy and lossless transport modes for RDMA, enabling smarter buffer management and reduced complexity.
• To advocate for integrated, specification-level congestion control and enhanced security in next-generation high-performance networking.
• To predict the convergence of HPC and datacenter networking through a unified, modernized Ethernet foundation that replaces TCP and RoCE within a decade.

Proposed method

• Analyzes the limitations of RoCEv2’s reliance on ECN-based congestion control, which uses a binary congestion flag and lacks fine-grained feedback, leading to slow convergence and poor performance under incast.
• Evaluates the inefficiency of PFC-based lossless networking in large-scale datacenters, where buffer overruns and congestion trees emerge due to complex interactions across layers.
• Proposes a new Ethernet-based RDMA stack that supports both lossy and lossless transport modes, decoupling buffer management from strict lossless requirements.
• Advocates for replacing go-back-n retransmission with selective acknowledgments to improve retransmission efficiency and reduce bandwidth waste.
• Introduces the need for in-network congestion signaling and direct switch-to-source feedback to reduce latency in congestion response and avoid receiver-bounce delays.
• Recommends designing protocols with extensibility for smart NICs, in-network computation, and secure memory semantics to support future workloads.

Experimental results

Research questions

• RQ1 Why does RoCE’s reliance on lossless Ethernet and ECN-based congestion control lead to poor performance and scalability in modern hyperscale datacenters?
• RQ2 How do higher link latencies in modern datacenter networks exacerbate incast problems and reduce the effectiveness of receiver-based congestion control?
• RQ3 What are the systemic drawbacks of using in-order delivery and go-back-n retransmission in high-throughput, high-message-rate workloads like AI training?
• RQ4 Can a unified, modernized Ethernet stack that supports both lossy and lossless modes improve network efficiency and reduce complexity compared to current RoCE and TCP/IP models?
• RQ5 How can next-generation RDMA networking better integrate with smart NICs, in-network telemetry, and security mechanisms to support future AI and HPC workloads?

Key findings

• RoCE’s ECN-based congestion control is ineffective due to its binary congestion flag, requiring many RTTs to converge, which leads to poor performance under incast and high-latency conditions.
• The use of PFC for lossless operation creates complex interactions across network layers, leading to congestion trees, victim flows, and inefficient buffer utilization in large-scale deployments.
• Higher end-to-end latencies increase the bandwidth-delay product, making even 1 MiB messages too small for effective incast handling via sender throttling, thus worsening incast issues.
• Legacy RDMA semantics such as in-order delivery and go-back-n retransmission severely limit network load balancing and prevent efficient multi-pathing, especially in long-lived AI training flows.
• The paper predicts that next-generation Ethernet will replace both RoCE and TCP/IP in hyperscale datacenters by supporting lossy and lossless modes, selective acknowledgments, and integrated congestion control.
• Modernized networking will enable smarter buffer management, reduce reliance on complex lossless fabric, and allow for more flexible, secure, and scalable high-performance networking for AI, HPC, and storage.