home *** CD-ROM | disk | FTP | other *** search
- Path: america.com!not-for-mail
- From: "Holger Kruse" <kruse@america.com>
- Newsgroups: comp.sys.amiga.networking
- Subject: Re: AmiTCP 4.0 and MTU=256 ???
- Date: Sat, 10 Feb 1996 07:54:42 -0500
- Organization: private Amiga site
- Message-ID: <kruse.53115025@america.com>
- References: <149214616750@Century.tng.oche.de> <9602082140.AA008tm@stokes.demon.co.uk>
- Reply-To: "kruse" <kruse@cs.ucf.edu>
- NNTP-Posting-Host: mco1-04.america.com
- X-NewsReader: IntuiNews 1.3a (7.9.95)
-
- In article <9602082140.AA008tm@stokes.demon.co.uk>, Philip Stokes
- <phil@stokes.demon.co.uk> wrote:
-
- > According to my TCP/IP books, The Internet Activities Board standards for
- > TCP/IP recommend that all networks should have an MTU of at least 576
- > octets. Certainly the AmiTCP installer will not allow you to set an
- > MTU of a lower value than this.
-
- You are misunderstanding the purpose of this requirement. All *networks*
- (i.e. *inner portions* of a TCP link, NOT the end point connections) are
- supposed to have an MTU of 576 or more.
-
- This is because most TCP/IP stacks do not implement MTU path discovery yet,
- so you want the lowest MTU on any portion of the network to be at least as
- high as the MTU on both end points of a TCP connection. Otherwise IP
- fragmentation will occur inside the network, which is Not A Good Thing,
- and can significantly increase the number of lost IP packets on the TCP
- link.
-
- To avoid IP fragmentation you would want the MTU on either end to be at
- least as high as the smallest UDP packet you are sending or receiving
- (including the IP overhead). If you use NFS across a SLIP/PPP link you
- probably do not want to go below 576. Otherwise it makes sense to lower
- the value to 296 (256 bytes payload plus 40 bytes IP header).
-
- There is no reason why the AmiTCP installer should prevent you from using
- low values like the above.
-
- > I don't know if the rhcslip.driver itself would support it but in view
- > of the above it is doubtful.
-
- There is no need for rhcslip.device (or any other driver, for that matter)
- to not allow small MTUs. If fact small MTUs can significantly improve
- interactive response time on slow serial links (because the average queueing
- delay is reduced) -- as long as you keep the MTU high enough that IP
- fragmentation does not occur.
-
- --
- Holger Kruse kruse@cs.ucf.edu kruse@america.com
- http://www.america.com/~kruse/home.html
-
-