Homa: A Receiver-Driven Low-Latency Transport Protocol Using Network Priorities (Complete Version) (1803.09615v2)

Abstract: Homa is a new transport protocol for datacenter networks. It provides exceptionally low latency, especially for workloads with a high volume of very short messages, and it also supports large messages and high network utilization. Homa uses in-network priority queues to ensure low latency for short messages; priority allocation is managed dynamically by each receiver and integrated with a receiver-driven flow control mechanism. Homa also uses controlled overcommitment of receiver downlinks to ensure efficient bandwidth utilization at high load. Our implementation of Homa delivers 99th percentile round-trip times less than 15{\mu}s for short messages on a 10 Gbps network running at 80% load. These latencies are almost 100x lower than the best published measurements of an implementation. In simulations, Homa's latency is roughly equal to pFabric and significantly better than pHost, PIAS, and NDP for almost all message sizes and workloads. Homa can also sustain higher network loads than pFabric, pHost, or PIAS.

• The paper introduces a receiver-driven framework that minimizes latency by dynamically managing network priorities, achieving 99th percentile round-trip times under 15 µs in a 10 Gbps network at 80% load.
• It employs a novel immediate transmission of a designated number of message bytes and dynamic priority allocation to mimic SRPT scheduling and reduce queuing delays.
• Performance evaluations demonstrate Homa’s superior handling of short and large messages compared to TCP-like protocols and alternatives such as pFabric, enhancing overall datacenter efficiency.

Homa: A Receiver-Driven Low-Latency Transport Protocol for Datacenter Networks

The paper explores Homa, a novel transport protocol specifically designed to address the unique needs of datacenter networks characterized by high volumes of short messages and the potential for very low latency communication. Unlike traditional transport protocols such as TCP, which are not optimized for these conditions, Homa leverages innovative mechanisms to closely approach the minimum latencies achievable by the underlying hardware.

Key Innovation and Design Principles

Homa introduces a receiver-driven architecture to manage and prioritize network flows. The central idea is to dynamically leverage in-network priority queues to ensure low latency, particularly for short messages. Each receiver in the network plays a pivotal role in allocating priority levels and managing flow control, contrasting traditional sender-driven approaches. This methodology addresses several critical challenges:

1. Blind Transmission of Short Messages: Homa allows a specified number of message bytes (RTTbytes) to be transmitted immediately without coordination to minimize latency, recognizing the limitation of packet scheduling for each message due to the time added by arbiter communication delays.
2. Dynamic Allocation of Priorities: Unlike static priority allocations, Homa dynamically adjusts priorities for each message at the receiver end. This dynamic approach mimics the shortest remaining processing time first (SRPT) scheduling policy, thereby minimizing queuing delays.
3. Controlled Overcommitment: To counteract bandwidth wastage associated with inactive sender responses, receivers are intentionally designed to overcommit by allowing several concurrent transmissions and using priority mechanisms to manage potential congestion.

Performance Evaluation and Comparative Analysis

Homa's performance metrics demonstrate significant improvements in both latency and throughput when compared to existing transport protocols:

• The implementation achieved 99th percentile round-trip times of less than 15 µs for short messages in a 10 Gbps network operating at 80% load.
• Despite being tailored for small message performance, Homa outperformed TCP-like approaches when handling larger messages due to its superior priority-based scheduling mechanism.

Simulations comparing Homa to protocols such as pFabric, pHost, and PIAS underscored Homa's proficiency in achieving low latency and effectively utilizing network bandwidth. Homa not only approximates near-optimal latency akin to pFabric's SRPT scheduling but also maintains higher network load sustaining capacities than these alternatives.

Implications and Future Directions

Homa represents a substantial advancement towards optimizing datacenter communication efficiency. Its architecture could redefine resource allocation and task scheduling strategies in latency-critical applications such as real-time analytics and distributed systems. Additionally, Homa's framework invites further exploration into enhancing hardware-level capabilities, such as supporting finer-grained priority queuing or packet-level preemption, to bridge the remaining latency gaps.

The paper's findings have practical ramifications in the design of future datacenter networks, potentially influencing TCP/IP stack implementations and necessitating further integration work to manage applications traditionally reliant on TCP's semantics. As Homa's operational paradigms gain traction, the protocol may spur ongoing investigations into more adaptive, receiver-driven network management techniques suited to evolving network environments.

In conclusion, the Homa protocol offers an insightful blueprint for reducing latency in datacenter networks through receiver-driven priority management and strategic overcommitment, setting a benchmark for forthcoming low-latency transport solutions.