MTU of the Internet?

William Allen Simpson wsimpson at greendragon.com
Sat Feb 7 17:46:01 UTC 1998


I hate to extend this thread any, but there is too much misinformation
to let slide....

> From: Stephen Sprunk <stephen at sprunk.org>
> Does anyone have data to show if any terminal servers or client stacks
> will honor TOS bits and/or known interactive port numbers when
> ordering packets for transmission across slow links?
>
Well, actually, every NAS for every vendor that I've ever worked with
honors the TOS bits (usually in numeric order) and gives precedence to
interactive traffic on outbound links, even fast ones, when there is a
queue.  Telebit, NEC, Rockwell, for examples.   Heck, that was even true
for the stack in a cellular phone (Qualcomm).

In response to the speculation that some bytes are forwarded without
determining TOS or packet size, I doubt that anybody is doing cut-thru
routing for 28.8 Kbps to 10 Mbps or vice versa.   Pretty silly idea.  No
gain.  An entire packet is sent at 10 Mbps in the time that it takes for
4 bytes to show up at 28.8 Kbps.

FYI, a larger number of these smaller MTU packets will actually cause
_more_ loss due to memory congestion.  Modern stacks do not allocate
packets on a byte by byte basis.  They allocate a buffer the size of the
largest allowable incoming packet on an interface (often called p-bufs).
So, a 40 byte packet takes the same memory as a 1500 byte packet.  You
don't get 3 times as many packets at 576.  There's no gain in trimming
the packet down after it comes in, the whole thing is just going to be
returned to the memory pool in milliseconds anyway.

As to the proposal that the big packets could be sliced and diced to
allow a smaller interactive packet thru -- well, the PPP fragmentation
header has been out there for years, but nobody really implements it,
even for multi-link.  Too much pain, not enough gain.  Same problems as
ATM.

The real problem is that too many OSes do not implement RFC-1123 Host
Requirements, even after 8 years.  As I remember the words of one
clueless product manager a year or so ago, listed in his byline as a
renowned Internet expert, complaining that (paraphrased) Simpson
believes that only those who have steeped in moldy RFCs should be
allowed to write network stacks.  It was pretty clear that he had not
bothered with reading too many of them on his way to becoming a
self-described expert, and very clear that his staff was none too
familiar with them, either.  And MicroSoft apparently has even more
problems than Apple.

The real answer for ISPs is to never ship Explorer on your configuration
CD, since it is impossible to make well-behaved.  Preconfigure Navigator
to 2 streams.  And even better, set Auto-Load Images Off.  There is a
nice big Images button to load the images on a page on those rare
occasions that you actually want to see them....

You will find that your users will be much happier with their networking
experience, and you will have fewer support calls.

WSimpson at UMich.edu
    Key fingerprint =  17 40 5E 67 15 6F 31 26  DD 0D B9 9B 6A 15 2C 32



More information about the NANOG mailing list