North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Re: Traceroute versus other performance measurement

  • From: smd
  • Date: Wed Nov 29 15:21:07 2000

| I thought the 5th tuple of 5-tuple is usually the L4 protocol (TCP, UDP,
| ICMP, etc.).  If your algorithm also takes DS values into account, you'd
| have a 6-tuple.

It's possible to take into account TCP/UDP/whatnot but breaking fate-sharing
based on something that is not inherently a hint to the routers of the Internet
seems rather strange, especially considering that you can end up in weird
situations where ICMP works and TCP doesn't, or vice-versa, making things
very hard to diagnose.

If a 5-tuple is being used, TOS/diffserv makes much more sense, because
the settor of that byte is explicitly signalling to the routing system
that packets are to be treated differently from others with different 
values in the same byte.

There is also the fact that splitting traffic across multiple paths
based on passenger transport protocol is unlikely to produce much of
a win anyway; the bulk of traffic by packets and by bytes is associated
with the other 4 elements is very likely to be TCP, and thus there is
not much better opportunity for statistical traffic separation than by
using the TOS/diffserv byte.  While most traffic simply is not coloured
today, the opportunity to colour in order to improve traffic separation
is greater when there is an element involved that you are allowed to change.

This is in essence a "poor man's explicit routing" in that the source
(or something in between) can move traffic between one source and
destination in a different way.

That is somewhat different from using L4 type, which is and must
always be invariant.