Network Performance Testing Summary

Network Performance Testing Summary

[John William]

This is a follow up message to the original "Abysmal Receive Performance" 
message. Thanks to everyone who e-mailed me with suggestions.

Well, after poking around, I eventually narrowed the problem down to the 
fact that the system BIOS did not enable PCI->RAM write posting. After I 
enabled that with the bridge optimization setting in 2.2.19, network RECV 
performance went from ~5Mbps to 41Mbps.

The curse of the HP Vectra XU 5/90 strikes again!

What is interesting is that I tried the NetGear FA310, FA311, 3COM 3cSOHO 
and 3C905C-TX cards and both the receive and transmit speeds (measured with 
both iperf and netperf) were so close to each other as to be a non-issue.

Several people e-mailed me to let me know that "card 'X' performance is 
terrible, I can only get good performance with card 'Y'". So, I just thought 
I should send this message out to set things a bit straight.

I know this machine isn't exactly cutting-edge, but I have to admit I was a 
little surprised to see all four of the cards I tested clustered so close 
together (so close, that you could call it 41.8Mpbs RECV and 52.2Mbps XMIT 
for all cards).

FWIW, it is somewhat interesting that 2.4.3 performance is so horrible with 
PCI->RAM posting disabled, whereas 2.2.19 doesn't take nearly the same hit. 
Using the 1.1.7 Tulip driver from Sourceforge didn't change this at all. 
With write posting enabled, things 2.4.3 ended up performing better than 
2.2.19.

So, thanks to everyone who helped me out. Looks like I now need to start a 
FAQ just for Vectra XU 5/90 owners so they can get their machines to work...

- John

_________________________________________________________________
Get your FREE download of MSN Explorer at http://explorer.msn.com

Re: Network Performance Testing Summary

[David Rees]

On Wed, Jun 06, 2001 at 02:52:03AM +0000, John William wrote:
> 
> The curse of the HP Vectra XU 5/90 strikes again!
> 
> What is interesting is that I tried the NetGear FA310, FA311, 3COM 3cSOHO 
> and 3C905C-TX cards and both the receive and transmit speeds (measured with 
> both iperf and netperf) were so close to each other as to be a non-issue.
> 
> Several people e-mailed me to let me know that "card 'X' performance is 
> terrible, I can only get good performance with card 'Y'". So, I just thought 
> I should send this message out to set things a bit straight.

Did you monitor CPU usage during these tests?

I did some throughput comparing a DLink RTL8139 based card to a 3C905C-TX card on a K6-2 450.  
Both managed to fully saturate 100Mbps.  However, the DLink used up ~90% CPU, and the 3Com 
only used about 50% CPU.  This was on 2.4.5, with the 8139too driver from 2.4.3.

-Dave