*BSD News Article 22213


Return to BSD News archive

Xref: sserve comp.os.386bsd.misc:1203 comp.os.linux:56062
Path: sserve!newshost.anu.edu.au!munnari.oz.au!constellation!osuunx.ucc.okstate.edu!moe.ksu.ksu.edu!vixen.cso.uiuc.edu!howland.reston.ans.net!agate!overload.lbl.gov!dog.ee.lbl.gov!horse.ee.lbl.gov!torek
From: torek@horse.ee.lbl.gov (Chris Torek)
Newsgroups: comp.os.386bsd.misc,comp.os.linux
Subject: Re: NetBSD TCP/IP network benchmarks
Date: 11 Oct 1993 10:05:27 GMT
Organization: Lawrence Berkeley Laboratory, Berkeley CA
Lines: 43
Message-ID: <34594@dog.ee.lbl.gov>
References: <CEnnD9.H8w@agora.rain.com> <1993Oct11.091056.7938@beaver.cs.washington.edu>
NNTP-Posting-Host: 128.3.112.15

In article <CEnnD9.H8w@agora.rain.com> davidg@agora.rain.com
(David Greenman) notes that
>>... You're testing the speed that you can write the shared memory on
>>the board - not the networking code. 

In article <1993Oct11.091056.7938@beaver.cs.washington.edu>
cmaeda@cs.washington.edu (Chris Maeda) writes:
>No, you're testing them all.  I was interested in comparing the
>performance differences due to operating system structure.  The five
>different configurations I measured had identical hardware, widely
>varying software OS structure, and widely varying performance.

Actually, you are both right.

Testing two (or more) different OSes for the same task on the same
hardware *does* test all the aspects of the software; but the nature
of the test affects *which* aspects figure the most strongly.

As it happens, Ethernet performance is, on reasonably fast CPUs with
anywhere near reasonably efficient networking code, a strong test of
the speed at which memory is copied and the number of copies
performed.  It is also a strong test of the care taken with the
particular Ethernet device driver involved.  Both of these factors show
up in the results, but in different ways.

As most modern machines can easily copy at greater-than-Ethernet
speeds, the actual throughput on such a test typically depends mainly
on the efficiency of the Ethernet driver.  Excessive memory-to-memory
copies show up only as increased CPU load.  That is, one system may
move 1.1 MB/s at 20% CPU load, while another may move the same 1.1 MB/s
at 80% CPU load, with the difference being due to the network system.
Contrariwise, one system might efficiently move 0.55 MB/s at 10% load
while another moves 1.1 MB/s but takes 90% of the CPU to do so.

(Note that I am not speaking specifically of any particular system, nor
even of PCs; this is a general `beware of what you are testing' sort
of comment.  I made up the various percentages above, for purposes of
illustration only.  Your mileage will vary.  Void where prohibited.
Taxes are the sole responsibility of recipient.  Warranty does not
cover deliberate misuse or acts of Goddess.)
-- 
In-Real-Life: Chris Torek, Lawrence Berkeley Lab CSE/EE (+1 510 486 5427)
Berkeley, CA		Domain:	torek@ee.lbl.gov