Cosa è successo
Stanford professor emeritus John Ousterhout is advocating for the adoption of Homa, a transport protocol designed to address issues in datacenters that he argues are poorly served by the Transmission Control Protocol (TCP). According to The Register, Homa is a message-based protocol that allows receivers to manage congestion control by prioritizing shorter messages using a shortest-remaining-processing-time (SRPT) . Ousterhout claims this approach reduces latency for short messages by an order of magnitude compared to TCP, citing a 99th percentile latency of 92 microseconds versus 1.2 milliseconds on a 100 Gbps network.
Homa was originally developed as part of a 2019 PhD dissertation by Behnam Montazeri, now a Google engineer. Ousterhout, who has retired from teaching, is now leading efforts to promote the protocol as a replacement for TCP in datacenter environments.
Unlike TCP, which is stream-based and lacks message-level visibility, Homa is message-based. It allows the receiver to manage congestion by knowing the size of incoming data from the first packet, enabling the receiver to schedule traffic and prioritize shorter, time-sensitive tasks.
Ousterhout claims that Homa can be installed as a Linux kernel module without requiring a system reboot and can operate alongside existing TCP applications, allowing for a gradual transition.
The protocol is currently undergoing standardization efforts through the IETF and is being integrated into the Linux kernel. It has already been backported to Red Hat Enterprise Linux 8 and 9.5.
Dettagli della fonte: theregister.com ↗
Perché è importante
AI workloads, particularly those involving large language models, require high-performance networking to handle weight gradients, model weights, and cache lookups. When network occurs, expensive GPU resources often sit idle, creating inefficiencies. Because TCP treats data as a continuous byte stream without inherent prioritization, it struggles to manage the mix of large data transfers and short, latency-sensitive control tasks common in modern AI infrastructure. Homa’s ability to explicitly schedule packets based on message length offers a potential technical solution to these bottlenecks, though it faces competition from other specialized protocols like RDMA and QUIC.
AI development relies on high-speed data movement for tasks like weight synchronization and lookups. Even millisecond-level delays in these transfers can cause significant underutilization of expensive GPU clusters.
TCP was designed for general-purpose internet traffic and lacks the granular control needed to distinguish between large background data transfers and small, urgent control messages. This leads to 'head-of-line blocking' and congestion issues that Homa aims to resolve.
The industry has previously turned to other solutions like RDMA, Fibre Channel, and Google’s QUIC to bypass TCP limitations. Homa represents a specific attempt to solve these issues within the datacenter by rethinking congestion control at the transport layer.
Meccanismo interattivo: come funziona realmente
Esplora la tecnologia alla base di questo sviluppo in modo interattivo.
Which component of an AI application is the machine-learning model itself?
Cosa guardare dopo
The Register reports that Ousterhout is currently drafting an IETF standardization document for Homa and working to upstream the protocol into the Linux kernel. While the protocol has been backported to Red Hat Enterprise Linux versions 8 and 9.5, its broader adoption remains unconfirmed. Observers should monitor whether Homa gains traction beyond prototype testing in financial services, as it faces skepticism from some network architects who question its necessity given existing alternatives like RDMA and AWS’s Scalable Reliable Datagram.
The primary hurdle for Homa is industry adoption. Network architect Ivan Pepelnjak has previously published critiques questioning the performance claims and the necessity of the protocol, suggesting it may be a 'solution looking for a problem.'
The success of Homa will depend on its ability to prove superior performance in real-world, large-scale AI deployments compared to established alternatives like RDMA or specialized cloud-native protocols.
Ousterhout is currently working with a large financial services firm on a prototype, which may serve as a test case for the protocol's viability in high-performance, -sensitive environments.