TCP's Retransmission Timer and the Minimum RTO

We study the performance impact of recently recommended TCP retransmission timer settings using a bare PC Web server with no operating system or kernel running in the machine. We first evaluate server performance in a test LAN with various settings of the alpha and beta constants used for computing SRTT and RTTVAR in the presence of varying levels of background traffic generated by conventional systems. We then study performance with different minimum RTO settings, and compare the performance of the bare PC Web server using the recommended timer settings with the performance of the Apache and IIS Web servers running on Linux and Windows respectively. We find that (1) no combinations of alpha and beta, or sampling strategies, perform consistently better than others under the different levels of background traffic; (2) lower minimum RTO settings than the recommended 1- second minimum will work when there is moderate background traffic, but the 1-second minimum is best when there are higher levels of congestion; and (3) using the standard timer settings but not using the TCP SACK option and congestion control mechanisms degrades bare server performance for some levels of background traffic.