[prev in list] [next in list] [prev in thread] [next in thread] 

List:       freebsd-hackers
Subject:    TCP time-outs on PPP line
From:       "Stefan Bethke" <stefan () Promo ! DE>
Date:       1996-07-01 20:10:41
[Download RAW message or body]

Hi,

we are connected through an analog leased line with two USR Courier, using 2.1.0R and \
pppd.

Since upgrading from Linux we experience some TCP "connection timed out" failures. It \
seems that at those moments someone is downloading from our HTTP or FTP server \
(looking at the modem leds and ftpwho/xferlog/http-log).

Trying to ping the terminal server of our ISP doesn't give a response; after 5 or so \
pings, ping says "No buffer space available".

My understanding is that the already established TCP connection quickly fills the \
output buffer and saturates the line; at this point, a packet inserted into the queue \
will be sent after (worst case) mtu / bandwith * queue length (ifq_maxlen) = 1500 / \
2880  * 50 = 27 seconds. (Or, taking fastq into account, 54 secs).

What's the solution? Possibilities are
- reduce ifq_maxlen, so ENOBUF will be returned earlier, so TCP will
  (hopefully) send slower earlier;
- reduce the mtu, so a full queue run will be faster, thereby allowing a new
  packet to be sent faster;
- devise a mechanism to promote those connections we want to precede our
  external users' connections to the fastq; possibly by promoting all
  packets smaller than xxx bytes to fastq;
- upgrade to a 2 Mbit line :-)

I didn't test any of these yet, because I rather like to have a slow but working \
line...

Any comments will be greatly appreciated.

Stefan Bethke

--
Promo Datentechnik      |  Tel. 040/431360-0
+ Systemberatung GmbH   |  Fax. 040/431360-60
Waterloohain 6-8        |  e-mail: stefan@Promo.DE
D-22769 Hamburg         |  http://www.Promo.DE/


[prev in list] [next in list] [prev in thread] [next in thread]