Architectural Implications of Performing Network Protocol Processing Closer to the Application
Matthew Faulkner, Matthew Jakeman and Stephen Pink
International Symposium on Performance Evaluation of Computer and Telecommunication Systems (SPECTS 2007)
San Diego, California (USA), July 16-18, 2007
SPECTS_Summary
It has been a while since high-speed host protocol processing has been a central concern of networking research. This paper expands on the ideas presented recently by Van Jacobson who, among others, played a large role in the development of efficient host processing almost two decades ago. Recently Jacobson has suggested moving the TCP/IP implementation out of the kernel of the operating system in multi-core CPU-based systems. This reduces the number of context switches and locks used, so that more concurrency, among other performance enhancements, are achieved. We do not attempt to modify this technique as the improvements are significant. Instead we attempt to identify and solve some of the problems inherent in changing the software architecture to achieve such a move. Taking the TCP/IP stack out of the kernel can have significant impact on the overall architectural aspects of Internetworking. This paper discusses this optimisation in relation to addressing, routing and the end-to-end principle of Internetworking. We also show that these changes can have a significant impact on applications. This paper raises a number of questions concerning user space TCP/IP and makes an attempt at some of their solutions.