Subject: Re: Performance issue in tcp_reass()
To: None <ragge@ludd.luth.se>
From: Thor Lancelot Simon <tls@rek.tjls.com>
List: tech-net
Date: 04/19/2004 18:45:14
On Mon, Apr 19, 2004 at 04:45:45PM +0200, ragge@ludd.luth.se wrote:
>
> In tcp_reass() the routine m_cat() is frequently called to append
> received mbufs to the end of a received chain. If the chain of
> received packets become long, the machine may become unusable for
> a long time if packets are lost. For example;
Speaking of packets being lost...
...if we aren't going to get a SACK implementation any time soon, I'd
like to turn on New Reno by default. The FTP server's been running
with it for a long time and seems to work OK. I'd also like to consider
implementing rate-halving; there's an old implementation for 1.3 floating
around which I might have time to bring up to date at some point after
the current academic year is over here in the U.S. (about three weeks
from now).
Opinions?
Thor