2002-10-17 08:54:31

by Roy Sigurd Karlsbakk

[permalink] [raw]
Subject: TCP Segmentation Offload (TSO) in 2.4?

hi

I've got some servers doing quite a lot of network traffic, and after reading
the article about TSO on kernel traffic
(http://www.kerneltrap.org/node.php?id=397), I wonder ... will this be
backported to 2.4? is it hard?

thanks

roy
--
Roy Sigurd Karlsbakk, Datavaktmester
ProntoTV AS - http://www.pronto.tv/
Tel: +47 9801 3356

Computers are like air conditioners.
They stop working when you open Windows.


2002-10-17 19:31:12

by David Miller

[permalink] [raw]
Subject: Re: TCP Segmentation Offload (TSO) in 2.4?

From: Roy Sigurd Karlsbakk <[email protected]>
Date: Thu, 17 Oct 2002 11:05:14 +0200

I've got some servers doing quite a lot of network traffic, and after reading
the article about TSO on kernel traffic
(http://www.kerneltrap.org/node.php?id=397), I wonder ... will this be
backported to 2.4? is it hard?

It isn't hard, but two issues:

1) You'll only get this with e1000 cards, and there were some
performance regression noted by some testers at IBM with
TSO enabled.

2) There are semantic issues still unresolved in the generic
TCP implementation parts, such as how to deal with the
fact that the way to get the most out of TSO is to
ignore the congestion window at connection startup. We
can only really make such a violation when talking to a
host on the local network etc. etc.

A 2.4.x backport is illadvisable until this is resolved.

2002-10-17 23:30:16

by Feldman, Scott

[permalink] [raw]
Subject: RE: TCP Segmentation Offload (TSO) in 2.4?

> It isn't hard, but two issues:
>
> 1) You'll only get this with e1000 cards, and there were some
> performance regression noted by some testers at IBM with
> TSO enabled.

Was this posted to the list? I remember Troy's results showing positive
results with TSO over SPECWeb.

-scott

2002-10-17 23:34:24

by David Miller

[permalink] [raw]
Subject: Re: TCP Segmentation Offload (TSO) in 2.4?

From: "Feldman, Scott" <[email protected]>
Date: Thu, 17 Oct 2002 16:35:18 -0700

> 1) You'll only get this with e1000 cards, and there were some
> performance regression noted by some testers at IBM with
> TSO enabled.

Was this posted to the list? I remember Troy's results showing positive
results with TSO over SPECWeb.

Mala Anand ([email protected]) has continually been disabling TSO in
2.5.x performance tests, at least this is how it has appeared to me.

Please, you two have a dialogue together and sort this out :-)