HOMA Protocol Aims to Replace TCP in AI Clusters

High latency, especially for small messages, results in significant waste of GPU resources as these powerful units spend considerable time waiting for synchronization. John Ousterhout confirmed this industry-wide struggle with latency through a live audience poll, indicating a pervasive issue across AI development.

Incast, where multiple nodes simultaneously transfer data to a single destination, is identified as a primary driver of network congestion. This phenomenon causes small messages to become trapped behind longer messages within switch queues. The resulting buffer exhaustion leads to dropped packets and necessitates costly retransmission cycles, exacerbating network inefficiencies.

Historically, network protocols, including TCP and RDMA, have delegated congestion management to the sender. Older methods relied on detecting packet loss to infer congestion, prompting senders to slow their transmission rates. More modern approaches utilize Explicit Congestion Notification (ECN) as a feedback mechanism, where switches mark packets to signal network pressure to the sender.

A top-of-rack switch, upon detecting that the queue length for an ingress port has reached a predefined threshold, begins marking all subsequent packets with an ECN tag. When these marked packets arrive at the receiver, the receiver extracts this ECN marking and notifies the original sender during subsequent acknowledgment transmissions.

Reactive congestion control mechanisms face significant instability due to information scarcity; senders only receive a single bit of congestion data. This control lag, requiring multiple round trips to adjust rates, is further complicated by multiple senders simultaneously attempting adjustments to the same destination. Such network volatility makes achieving stable convergence of transmission rates nearly impossible.

TCP's fundamental limitation lies in its sender-side approach to congestion control, which often results in constant oscillation between sending too much and too little data. This inefficiency leads to significant queue buildup within the network, a problem that has been recognized in research for over two decades. Congestion is only detected after delays have already occurred, making proactive management difficult.

TCP and RDMA protocols operate on a continuous byte-stream model, which lacks inherent differentiation between distinct messages. This design prevents senders from effectively prioritizing short messages. Consequently, head-of-line blocking occurs, where smaller, time-sensitive messages get stuck behind larger, less urgent data flows, leading to significant inefficiencies in high-performance computing environments.

HOMA is specifically engineered for data center environments and diverges from TCP and RDMA in virtually every major design decision. Its core focus is on effectively managing a diverse mix of large and small messages, a critical requirement for modern AI clusters.

HOMA originated from Benam Montazeri's PhD dissertation and has been implemented as a Linux kernel module. Ousterhout is currently working through the process of getting it upstreamed into the kernel, and it is readily available on GitHub for download and experimentation.

HOMA's architecture is message-based rather than stream-based, with the fundamental unit being a Remote Procedure Call (RPC). This design allows HOMA to transport message length information all the way down to the transport layer. By knowing message lengths, HOMA can predict data volume upon receiving the first packet and use Shortest Remaining Processing Time first (SRPT) scheduling to prioritize short messages effectively.

A key difference in HOMA is its receiver-driven approach to congestion control. Congestion primarily occurs at the receiver's downlink, giving the receiver complete and superior information about incoming data volume. This comprehensive knowledge allows HOMA to respond faster and more precisely to congestion events, optimizing network flow more effectively than sender-driven methods.

HOMA employs a credit-based grant system where the sender splits messages into initial 'unscheduled' packets and subsequent 'scheduled' packets. The receiver sends grant packets to trigger the transmission of this scheduled data. This mechanism allows receivers to strategically delay grants, preventing congestion before it occurs, and naturally prioritizes shorter messages to implement SRPT logic effectively.

HOMA leverages modern data center switches, which typically feature multiple egress queues, often eight, to dynamically prioritize traffic. It assigns messages to these queues based on packet headers, ensuring that shorter messages are pushed to higher-priority queues. This capability allows short, time-sensitive messages to preempt longer data flows, effectively reducing head-of-line blocking that plagues traditional protocols.

Benchmarking message sizes from 50 bytes to 1 megabyte reveals HOMA's superior performance over TCP. For short messages, HOMA achieves a P99 tail latency that is 13 times faster than TCP, significantly reducing critical delays. Surprisingly, HOMA also improves performance for long messages, nearly doubling the speed compared to TCP. This demonstrates that HOMA's 'run-to-completion' approach surpasses TCP's fair scheduling, which often causes short messages to be delayed by longer ones, regardless of priority.

Ousterhout emphasizes the growing importance of short messages in AI clusters and encourages engineers to evaluate if short-message latency is a bottleneck in their systems. He is dedicating his current career to HOMA's development and adoption, inviting interested parties to reach out for collaboration, debugging, and technical support to ensure successful implementation.