Tuning Gigabit

[Martin Nilsson]

David Gilbert wrote:
> Well... I don't have any disk controllers to test, but I've verified
> this behaviour with em, bge, sk, and nge chipsets.  Of these, as I've
> said before, the em shines.  

I have made some simple speed testing of gigabit NICs too. My results 
are that Linux 2.4.18+ is usually faster and have much lower CPU load 
than FreeBSD 4.8. I have tested the following NIC:s em (1000/XT,Anvik), 
nge(32/33), bge (Altima 1002), ti (Netgear GA620T).
For FreeBSD I've found that use nothing beats the ti with its special 
firmware.

> The bge has good performance, too ... but
> we've seen a lot of corrupted routed packets ... it has some
> interaction with some motherboard chipsets.

The VIA chipsets often have very poor PCI bus performance if you don't 
load the 4 in 1 drivers (windows only). The max PCI speed I've been able 
to get with (semi old ~1.5years) VIA chipsets are about 70MB/s measured 
with a SCSI disk array on a 32/33 bus. The intel, SiS and serverworks 
chipsets perform as expected, I haven't tested any Nvida boards.

> It's in the back of my mind.  I don't think I'll have time for this
> BSDCon, but maybe soon thereafter.  It's getting to the point where we
> should have a BSD Journal.

I would be interested in helping to test this, I have a bunch of lab 
boxes with 64/66PCI or PCI-X slots and NICs to work with. I'm able to 
get access to most new motherboards and chipsets to test chipset 
specific performance. What I need is someone to help me develop useful 
things to test, validate results and to check testing methodology.


	/Martin
-- 
Martin Nilsson M.Sc CS&E, CTO Mullet Scandinavia AB, Malmö, SWEDEN
E-mail: martin at mullet.se, Phone: +46-(0)708-606170, Web: www.mullet.se

Our business is well enigineered servers optimised for FreeBSD & Linux