Subject: Re: Intel i82547 performance problems in wm(4)
To: None <tech-net@NetBSD.org>
From: Matthias Scheler <firstname.lastname@example.org>
Date: 07/16/2004 12:25:24
On Fri, Jul 16, 2004 at 12:14:58PM +0200, Mipam wrote:
> > Why not? The card will calculate the correct checksum whether the firewall
> > manipulated the packet or not. The firewall only has to cope with the
> > fact that it will see outgoing packets with incorrect checksums.
> But shouldnt tcp packets with incorrect tcp checksums be discarded
> according rfc 793, causing a firewall to drop such a packet?
1.) When the packets hits the wire the checksum will be correct.
2.) Packets with incorrect checksums on the send queue have a marked
in the data structure so that the firewall software can recognize
them and neither perform an initial checksum check nor bother to
correct the checksum after changing the packet.
> And as for the benefits of offloading engines, on local networks with
> several cards which support offloading it can make a difference (large
> frames (if the switch supports them :-)), less interrupts) and so large
> tcp windows can make a difference.
Offload support in NetBSD doesn't affect TCP window handling. You are
probably refering to LSO (large segment offload) which NetBSD doesn't
support yet. NetBSD supports offloading IP, TCP and UDP checksums
to the card. And that simply saves CPU cycles on the local machine.
While the amount of saved CPU cycles per time interval of course
depends on the utilized bandwidth it works for LAN and WAN connections.
> On the internet jumbo frames dont work, ....
Jump frames are a different concept than checksum offloading.
Matthias Scheler http://scheler.de/~matthias/