TransBridge: A Transparent Communication Middleware with Unified RoCE and TCP Semantics

Area of Science:

• Computer Science
• Network Engineering
• Distributed Systems

Background:

• Low-latency communication is critical for edge intelligence applications like autonomous driving.
• Traditional TCP protocol overhead hinders performance in these scenarios.
• High-performance RoCE protocol is difficult to integrate with existing Socket APIs.

Purpose of the Study:

• To develop a communication middleware that transparently bridges TCP and RoCE protocols.
• To enable unmodified Socket-based applications to leverage RoCE performance.
• To address stringent communication latency requirements in edge intelligence.

Main Methods:

• Proposed TransBridge, a lightweight user-space communication middleware.
• Implemented a transparent user-space compatibility architecture.
• Developed a microsecond-level low-latency transmission engine bypassing kernel overhead.
• Utilized a decentralized peer-to-peer architecture with deferred buffer updates for resource management.

Main Results:

• Achieved 5.926 μs average round-trip latency for 16 B messages.
• Reached 20.254 Gbps throughput for 16 KB messages on a RoCE network.
• In Fast DDS evaluation, achieved 188 Mbps throughput and ~150 μs latency.

Conclusions:

• TransBridge provides transparent and effective RoCE acceleration for Socket-based applications.
• The middleware significantly outperforms mainstream schemes in edge environments.
• Enables high-performance communication essential for demanding edge intelligence tasks.