Date: Wed, 04 Oct 1995 23:42:23 +0100 From: Andras Olah <olah@cs.utwente.nl> To: current@freebsd.org Subject: Re: lmbench 1.1.5 vs 2.2-current Message-ID: <24072.812846543@utis156.cs.utwente.nl> In-Reply-To: Your message of "Wed, 04 Oct 1995 13:00:25 %2B1000." <199510040300.NAA09750@godzilla.zeta.org.au>
next in thread | previous in thread | raw e-mail | index | archive | help
A few comments on the TCP lmbench results: In my view, the slight increase in TCP latency is partly due to the increased processing required by T/TCP. I haven't done any tests, neither did I count the extra instructions, so I cannot give a specific figure of how much is this overhead. I don't think it's significant. As for the throughput, I collected a trace with tcpdump. It turned out that the 0.21MB/s throughput is caused by an interference of the 16384 byte MTU of the loopback i/f and the Nagle algorithm. The transfer was running in a lock-step, when a few segments were sent and then the sender was waiting for a delayed ack (~200ms). This situation is described in a recent IEEE/ACM Trans. on Networking in the context of IP over ATM. Using the TCP_NOPUSH socket option increased the bw to 2.9MB/s on my noname 486/40MHz. I guess TCP_NODELAY would do the same. Andras P.S: I don't have lmbench at home (where I did the test), but a program bw_tcp by L.M. from 1994 which I believe is a precursor of the lmbench package.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?24072.812846543>