On Wed, Apr 03, 2002 at 07:07:50PM -0500, Kunal Trivedi wrote:

> Lennert:

Kunal,


> > Did you just look at 'top' for determining this?  Time spent in
> > interrupt context isn't counted in that figure :-)
>
>  I used 'vmstat'.

OK.. that isn't very reliable for this workload.


> Here, are some results. I used 'iperf' and 'nttcp' both.
> Before, you read further in details, I got performance of
> around 520-580 Mbits/sec both in UDP/TCP (using both iperf
> and ttcp). And one of my end machines is running FreeBSD 4.5

If you even get this without using bridging (end-to-end),
or if you also get this with routing, I guess that you
will have a much higher chance of finding interesting
conversation on either the linux kernel list, or the
netdev list ([EMAIL PROTECTED]).


> Questions:
> 
> 1) Why br-machine (running bridge) cpu load went high during TCP (upto
> 44%) and remained low(upto 18%) during UDP ?

Do you have firewalling modules loaded on the bridge?  Do you
have connection tracking loaded?  TCP connection tracking I
think is heavier than UDP connection tracking.


> 2) Why Linux machine could not push more than 580 Mbits/sec.
> (even i gave iperf -c -u -b 900 m). While FreeBSD could push
> 950 Mbits/sec. (BEWARE: Here I am not saying that Linux is bad
> or whatever...). Is it a driver issue ?

This could be a lot of things, but if FreeBSD has better
performance on the same hardware, it's likely to be a driver
issue.  Do you have other cards available for testing?


> 3) Currently i am only doing bridging, but i am planning to
> do more on top of bridging (e.g., shaping, policing based on
> different attributes). So, pkt might goto Layer 3 in that case.
> Will performance degrade in that case ?

I think routing is definitely heavier than bridging, but I
would be interested to see your performance figures for
routing.


> 4) Lennert:
> 
>       > Yeah, odd.  I get full wire speed (120 megabytes/sec) over TCP
>       > between a dual Athlon 1.2ghz and a Pentium III 1.2ghz (where
>       > the athlon is the transmitting host, network card interrupt
>       > bound to CPU#0).
>
> If it is possible, could you tell me what else did you use  ?
> (Specification of machine/card etc., application which measured
> performance etc.)

The dual Athlon has a Tyan Thunder K7 mainboard, with two Athlon
MP 1200MHz processors and 1GB of DDR RAM.  Used kernel is
2.4.19-pre2-ac2.  Used network card is (I think) the 3com 3C996B,
using the version 0.96 tg3 driver.  The card is in a 33MHz 64-bit
PCI slot.

The pentium III 1.2 has an intel STL2 server mainboard (dual
processor capable), with only one CPU plugged in.  It runs a
uniprocessor 2.4.19-pre2-ac2 kernel.  It has the same network
card as the athlon, and the same driver.

Neither of the machines have netfilter modules loaded during
the test (but both kernels were compiled with netfilter
support as modules).

Measuring application is ttcp.  No zero-copy was used, and no
animals were harmed in the collection of these figures.


cheers,
Lennert
_______________________________________________
Bridge mailing list
[EMAIL PROTECTED]
http://www.math.leidenuniv.nl/mailman/listinfo/bridge

Reply via email to