*BSD News Article 43180


Return to BSD News archive

Xref: sserve comp.os.386bsd.questions:16920 comp.os.386bsd.development:3274
Path: sserve!newshost.anu.edu.au!harbinger.cc.monash.edu.au!newsserver.trl.OZ.AU!huon!lukem
From: lukem@melb.cpr.itg.telecom.com.au (Luke Mewburn)
Newsgroups: comp.os.386bsd.questions,comp.os.386bsd.development
Subject: Re: ethernet throughput
Date: 5 Mar 95 22:10:19 GMT
Organization: CPR Project, ITG, Telstra
Lines: 66
Message-ID: <lukem.794441419@huon>
References: <3it7m8$o39@clavin.uprc.com>
Reply-To: lm@cpr.itg.telecom.com.au
NNTP-Posting-Host: 144.136.63.213
Keywords: ethernet
X-Newsreader: NN version 6.5.0 #2 (NOV)

z056716@uprc.com (LaCoursiere J. D. (Jeff)) writes:

> What kind of throughput are people getting out of high-end ethernet cards
> and FreeBSD?  I was doing ftp's between two Sparc LX boxes on a subnet by
> themselves and was seeing about 4Mb transfer rates.  Does PC hardware exist
> that could reach these speeds?  I seem to remember some quotes of 900Kb
> with 32 bit cards, which makes me skeptical.  I only get 400Kb through my
> 16bit WD8003 cards (486DX2/66 FreeBSD1.1.5.1 to 386DX/40 FreeBSD 1.1.5.1).

> If you have reached 4Mb, please email me your motherboard type and ethernet
> card!

What do you mean by "4Mb"? 4Mega-bits/second (500 kilobytes/second?)

I've gotten >840 Kbyte/second read _and_ write over NFS between two
NetBSD-current boxes.

Configuration:

Server: 486dx33, 8MB RAM, ASUS SP3G (PCI/ISA) motherboard, SMC 8216C
ethernet card, default buffer cache (~600K). 2 x DEC RZ26 1GB hard drives.

Client: 486dx2-66, 8MB RAM, cheapo motherboard, SMC 8216C ethernet
card, default buffer cache (~600K). Acting as diskless client to
server, with local swap partition.

Server has NFS code to do asynchronous writes (remove IO_SYNC from
a 'ioflags' variable initialization in /sys/nfs/nfs_serv.c). This is
not recommended for important commercial applications, but when the
client & the server both sit under my desk, I figure I'd lose if the
server crashed anyway, synchronous writes over nfs from the client,
or asynchronous writes...

Here's the results of 'bonnie', a reasonably good disk benchmark.
             -------Sequential Output-------- ---Sequential Input-- --Random--
             -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine    MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU  /sec %CPU
nfs:/z/1   16   675 60.1   848 19.9   221 13.4   644 58.3   863 19.1  44.2 11.8
local:/z/1 16   802 97.8   821  7.4   656 19.5   855 99.0  2205 32.5  61.6  7.2


There was slightly more CPU load (other tasks) on the server ('local')
when its tests were occuring. 

Note that other tests (dd if=/dev/null of=some_file_on_nfs_mounted_fs bs=xk,
where 1 < x < 128) shows similar results.


Read what you will into my results, but I can think of three factors
that attribute to this:
- the ASUS SP3G motherboard in the server, with the high speed SCSI
  interface to the drives
- the SMC 8216C elite ultras in the server & client
- the async NFS hacks (these will only speed up NFS of course, I
  haven't benchmarked ftp yet)

I heartily recommend the ethernet cards and definately the
motherboard. NetBSD & FreeBSD current (and possibly the newest
'offical releases') support the NCR 53c810 SCSI chip on the SP3G).


--
                                                                  Luke Mewburn
                                                              <lm@werj.com.au>
``Think of it as Evolution in Action''
    -- 'Oath of Fealty', Niven & Pournelle