Implementation and Analysis of TCP/IP Offload Engine and RDMA Transfer Mechanisms on an Embedded System

摘要

The speed of present-day network technology exceeds a gigabit and is developing rapidly. When using TCP/IP in these high-speed networks, a high load is incurred in processing TCP/IP protocol in a host CPU. To solve this problem, research has been carried out into TCP/IP Offload Engine (TOE) and Remote Direct Memory Access (RDMA). The TOE processes TCP/IP on a network adapter instead of using a host CPU; this reduces the processing burden on the host CPU, and RDMA eliminates any copy overhead of incoming data streams by allowing incoming data packets to be placed directly into the correct destination memory location. We have implemented the TOE and RDMA transfer mechanisms on an embedded system. The experimental results show that TOE and RDMA on an embedded system have considerable latencies despite of their advantages in reducing CPU utilization and data copy on the receiver side. An analysis of the experimental results and a method to overcome the high latencies of TOE and RDMA transfer mechanisms are presented.