Improving Processing Performance of Linux TCP SACK Implementation

Linux TCP implementation has significant limitations in loss recovery performance due to Selective Acknowledgment (SACK) related processing latencies when the outstanding window is large. In this paper we address a number of these shortcomings by extending earlier work by others using fast access shortcut variables to the expected next-tobe-processed segment and through novel resegmentation of SACKed octets that form continuous areas. We show that performance of the SACK scoreboard related operations do not scale without our modification whereas Linux kernel implementing the proposed modifications is able to perform at 1GbE and beyond speeds. In most of the cases ACK processing times are cut very dramatically, the measured results show at least 90% decrease in the ACK processing times.