Path: utzoo!utgpu!news-server.csri.toronto.edu!eecg.toronto.edu!leblanc
Newsgroups: comp.sys.amiga
From: leblanc@eecg.toronto.edu (Marcel LeBlanc)
Subject: Re: A3000UX Seems Fated (Kill file alert!)
Message-ID: <1991Jan9.151718.5042@jarvis.csri.toronto.edu>
References: <1990Dec25.045537.13517@NCoast.ORG> <1990Dec25.234322.836@ddsw1.MCS.COM> <17091@cbmvax.commodore.com> <1991Jan07.054805.13284@ddsw1.MCS.COM> <17197@cbmvax.commodore.com>
Date: 9 Jan 91 20:17:19 GMT
Lines: 42

daveh@cbmvax.commodore.com (Dave Haynie) writes:

>In article <1991Jan07.054805.13284@ddsw1.MCS.COM> karl@ddsw1.MCS.COM (Karl Denninger) writes:
...
>>RS/6000s use 3Com boards for Ethernet the last time I looked.  3C523s.  
>>They are DOGS.  I have worked with one, and it did something like 
>>150KB/second of real throughput doing real work (ie: serving up an FTP 
>>file).  That's terrible!  Sorry, but 15% of the cable bandwidth doesn't cut
>>it folks!

Then show us a machine that can do better.  (Relax, the LAST thing I want to
do is defend IBM or the RS/6000.)

>As I recall, the Personal Workstation folks seemed quite impressed by the
>I/O capabilities, if not the CPU speed claims, on the RS/6000.  Its SCSI,
>on the MCA bus, was one of the fastest they've tested.  I thought the Ethernet
>was too.  It depends alot on your Ethernet load, too.  We have days around
>here where the VAXen and big MIPS machines on our network have trouble
>hitting 15KB/second in ftp.
>...  You don't need much CPU power to have a fast network,
>just a good network design and a good network implementation.  

Having spent the last 2 years researching ways of improving the speed of
network interfaces, I can safely say that implementation is extremely
important.  150KByte/s at the application level is really very good.  You
are paying a performance penalty because of the fact you are running UNIX.
Incoming packets are copied multiple time as they go from the Ethernet
hardware to kernel buffers, and eventually to user space.  The limiting
factor with the current (poor) BSD TCP/IP implementation is most often
memory speed (especially in this case, where maximum packet size will be
used).  I doubt that IBM has spent much time optimizing their implementation
since researchers have only come to realize in the past few years where the
real protocol processing bottlenecks are.  Only now that people are locating
the real bottlenecks can _great_ network interfaces be designed.

>>Karl Denninger (karl@ddsw1.MCS.COM, <well-connected>!ddsw1!karl)

>Dave Haynie Commodore-Amiga (Amiga 3000) "The Crew That Never Rests"

Marcel A. LeBlanc  --  Electrical Eng. Computer Group, Univ. of Toronto
-----------------------------------------------------------------------
leblanc@eecg.toronto.edu		else: uunet!utcsri!eecg!leblanc