Problems with network performance.

Donald Becker
Mon Apr 26 13:39:45 1999

On Mon, 26 Apr 1999, Mike Brodbelt wrote:

> I've recently been investigating network performance with Linux boxes on
> However, there is clearly something wrong with the machines, and I'm
> boxes are using a Tulip chipset card, and I'm hoping someone here can
> Normally, I expect 900 kb/s upwards on a quiet 10 Mb coaxial segment,
> onto, using Samba. I tend to think that the error rate is way too high,
> but I don't know what the cause could be. The network is 10base2, in a

> eth0      Link encap:Ethernet  HWaddr 00:00:C0:0C:75:E8
>           RX packets:18457653 errors:1717 dropped:0 overruns:0
>           TX packets:15145268 errors:79182 dropped:0 overruns:0

You should be reporting the statistics from /proc/net/dev, rather than
summarized numbers from 'ifconfig'.  But I can tell immediately that
something is broken, almost certainly your 10base2 cabling.

>           RX packets:18831732 errors:1777 dropped:0 overruns:0
>           TX packets:15463638 errors:81207 dropped:0 overruns:0
> Errors shot up to 81K during mornings testing.

Ouch!  Try repeating the test using a dedicated 10base2 network, with no
breaks in the cable.

> Machine specifications:-
> A - Intel PII-350, 256Mb RAM, 12GB SCSI disk, RAID 5, SMC EtherEZ NIC,
> with DEC 21041 tulip chipset (tulip.c v0.89H). Linux 2.0.36, RedHat 5.1,
> Samba 2.0.3
> B - Intel P-166, 32Mb RAM, 2GB SCSI disk, DEC DS21041 Tulip (tulip.c
> v0.79). Linux 2.0.32, RedHat 4.1, Samba 1.9.17

Updating the driver will likely *not* change the numbers.
Almost all Tulip driver updates are for media selection code.
The only operational changes for the 21041 are
  Dynamic Tx threshold adjustment.
  Increased PCI burst transfer settings
    (I removed the work-around for broken '486 motherboards)
I doubt that either is an issue at 10mbps.

Donald Becker
USRA-CESDIS, Center of Excellence in Space Data and Information Sciences.
Code 930.5, Goddard Space Flight Center,  Greenbelt, MD.  20771