Return to BSD News archive
Xref: sserve comp.os.386bsd.misc:1214 comp.os.linux:56125 Newsgroups: comp.os.386bsd.misc,comp.os.linux Path: sserve!newshost.anu.edu.au!munnari.oz.au!news.Hawaii.Edu!ames!agate!howland.reston.ans.net!europa.eng.gtefsd.com!uunet!psgrain!ee.und.ac.za!hippo.ru.ac.za!kudu!g89r4222 From: g89r4222@kudu.ru.ac.za (Geoff Rehmet) Subject: Re: NetBSD TCP/IP network benchmarks Message-ID: <g89r4222.750411858@kudu> Sender: news@hippo.ru.ac.za (Usenet News Admin) Organization: Rhodes University, Grahamstown, South Africa References: <CEnnD9.H8w@agora.rain.com> Date: Tue, 12 Oct 1993 07:44:18 GMT Lines: 78 In <CEnnD9.H8w@agora.rain.com> davidg@agora.rain.com (David Greenman) writes: >>UDP round trip latency (in milliseconds) >> >>system message size (bytes) >> 1 100 512 1024 1472 >>------------------------------------------------------------- >>NetBSD 0.8 2.63 3.49 6.04 9.54 12.50 >>Mach 2.5 1.83 2.44 5.19 8.51 11.41 >>Mach 3.0(UX) 3.96 4.67 7.86 11.65 15.00 >>Mach 3.0(BSDSS) 4.64 5.37 8.95 13.23 16.84 >>Mach 3.0(library) 2.12 2.68 5.41 8.74 11.66 > I just grabbed your latency benchmark program. Here is the >result (again: client=486DX2/66 w/8013, server=486DX/33 w/8013): >FreeBSD-1.0E 1.0011 1.356 2.7415 4.4723 5.9395 >...and I just read your message about your board being an 8bit 3c503. >First, I'm sure you agree that the 1 byte ping test above isn't going >to be effected be the type of card that is being used, and in this case >the software is identical (the 'ed' driver supports the 80x3, 8/16bit >3c503, NE1000 and NE2000). > Now, let me say that attempting to test network performance by using >an 8bit ethernet card is rediculous. You're testing the speed that you >can write the shared memory on the board - not the networking code. A >test with the client/server being localhost would be more telling: I was running some tests of the Linda system which I am currently developing last night. The tests showed a quite significant difference in the amount of throughput for machines using 16bit ethernet cards (as opposed to 8bit). Roughly speaking, for the same amount of CPU time used, the machines with 16bit cards processed about double the number of evals than the machines with 8 bit cards did. (The program being run was an n-queens program, with board sizes varying from 12 to 15. The tuple space server was running on a Sun10/52.) The FreeBSD systems being used are 486DX33's with 8/16M RAM, and either WD8003E 8 bit or SMC Elite 16 bit cards. Other systems being used during the runs were various Sun workstations. If anyone is interested I can send them some of my performance results - they aren't really all that interesting. Here is the breakdown of the performance of the FreeBSD Systems: Notice that the machines with 16bit cards process about double the number of evals per second. The example given here is for a 15x15 chess board, with workers being spawned at a depth of 4. The real time taken for the test was 618 seconds. Machine CPU sec Evals Evals/sec NIC RAM ------- ------- ----- --------- ---- --- braae.ru.ac.za 569.29 630 1.1066 16bit 16M csmsc1.ru.ac.za 515.25 563 1.0926 16bit 16M csrbh08.ru.ac.za 593.01 309 0.5210 8bit 8M csrbg02.ru.ac.za 592.10 306 0.5168 8bit 8M csbodo2.ru.ac.za 593.02 306 0.5160 8bit 8M csds.ru.ac.za 570.42 639 1.1202 16bit 8M (just for fun ...) omega.ru.ac.za 952.13 3733 3.9206 Sun10/52 All the 486's are running FreeBSD 1.0EPSILON. I know this is not a direct benchmark of network performance, but it does illustrate very graphically the difference between using 8 and 16bit ethernet cards. (Oh also - the tests were run between 12pm and 1am - giving a low overall network load). Geoff. -- ===========================csgr@alpha.ru.ac.za================================ Geoff Rehmet, Parallel Processing Group, | ____ _ o /\ Computer Science Department, | ___ _-\_<, /\/\/\ Rhodes University, RSA. | (*)/'(*) /\/\/\/\/\