Optimizing TCP Receive Performance

The performance of receive side TCP processing has traditionally been dominated by the cost of the `per-byte' operations, such as data copying and checksumming. We show that architectural trends in modern processors, in particular aggressive prefetching, have resulted in a fundamental shift in the relative overheads of per-byte and per-packet operations in TCP receive processing, making per-packet operations the dominant source of overhead. Motivated by this architectural trend, we present two optimizations, receive aggregation and acknowledgment offload, that improve the receive side TCP performance by reducing the number of packets that need to be processed by the TCP/IP stack. Our optimizations are similar in spirit to the use of TCP Segment Offload (TSO) for improving transmit side performance, but without need for hardware support. With these optimizations, we demonstrate performance improvements of 45-67% for receive processing in native Linux, and of 86% for receive proces...

• Type: Conference
• Year: 2008
• Where: USENIX
• Authors: Aravind Menon, Willy Zwaenepoel
• Keywords: Architectural Trend, Operating System, Receive Processing, TCP Receive Processing, USENIX 2008