From owner-freebsd-current Wed Dec 9 14:31:54 1998 Return-Path: Received: (from majordom@localhost) by hub.freebsd.org (8.8.8/8.8.8) id OAA06433 for freebsd-current-outgoing; Wed, 9 Dec 1998 14:31:54 -0800 (PST) (envelope-from owner-freebsd-current@FreeBSD.ORG) Received: from duke.cs.duke.edu (duke.cs.duke.edu [152.3.140.1]) by hub.freebsd.org (8.8.8/8.8.8) with ESMTP id OAA06419 for ; Wed, 9 Dec 1998 14:31:51 -0800 (PST) (envelope-from gallatin@cs.duke.edu) Received: from grasshopper.cs.duke.edu (grasshopper.cs.duke.edu [152.3.145.30]) by duke.cs.duke.edu (8.9.1/8.9.1) with ESMTP id RAA15935 for ; Wed, 9 Dec 1998 17:31:45 -0500 (EST) Received: (from gallatin@localhost) by grasshopper.cs.duke.edu (8.9.1/8.9.1) id RAA86198; Wed, 9 Dec 1998 17:31:37 -0500 (EST) (envelope-from gallatin@cs.duke.edu) From: Andrew Gallatin MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Date: Wed, 9 Dec 1998 17:31:36 -0500 (EST) To: freebsd-current@FreeBSD.ORG Subject: increased UDP latencies in -current X-Mailer: VM 6.43 under 20.4 "Emerald" XEmacs Lucid Message-ID: <13934.54777.633326.333061@grasshopper.cs.duke.edu> Sender: owner-freebsd-current@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.ORG In the process of doing some Myrinet benchmarks, I noticed that UDP latencies have increased markedly after upgrading from 2.2.5-stable to 3.0-current. On identical machines (300Mhz PII, 440lx chipset, same memory configuration) a kernel built from circa Jan. 18, 1998 2.2-stable sources yields UDP small message latencies about 10% better than a 3.0-current kernel using both the netperf UDP_RR test & HBench-OS's lat_udp. A 3.0 kernel from mid July yields slightly better results than a recent -current, but not by much. Tests run across the loopback interface for a 1 byte message: RR/sec via us/msg via OS rev netperf(1) HBench-OS(2) --------- -------- ---------- 2.2.5 15502.17 56.6806 3.0 (July 7) 13787.90 64.9367 3.0 (Dec 1) 13656.60 66.5090 (1)netperf2.1pl3, run via netperf -tUDP_RR -l60 -i 10,2 -Hlocalhost (2)hbench: lat_udp.c,v 1.4 1997/06/27 00:33:58 abrown Exp $, run as lat_udp 10000 localhost Running tests between machines with real hardware (full duplex 100Mb ethernet over the fxp interface, as well as Myrinet/Trapeze) mirrors these differences. Larger messages (tried 1k and 8k) also mirror these results (2.2.5 is ~10us faster). TCP latencies appear to be nearly identical between OS revisions, as do UDP and TCP bandwidths. Does anybody know where the increased UDP latency is coming from? I initially thought the culprit might be no longer moving pcb's to the head of the hash chain in in_pcblookup_hash (removed in in_pcb.c,1.38). I tried replacing this code, but it doesn't make much of a difference... I suppose its not terribly important in the scheme of things, but it would be interesting to know why things are slower. ------------------------------------------------------------------------------ Andrew Gallatin, Sr Systems Programmer http://www.cs.duke.edu/~gallatin Duke University Email: gallatin@cs.duke.edu Department of Computer Science Phone: (919) 660-6590 To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message