Re: RFC2873 manually setting precedence.

[Alan Hannan <[e-mail]>]

  Phillip,

> Since most DSCP and QoS functions are normally handled within a single
> network domain, what impact would it have to manually set precedence at the
> edge.
> 
> for example,
> CE = customer edge, or host systems (TCP endpoints)
> PE = Provider edge,
> P   = Provider Core, 
> 
> CE - PE - P - PE - CE

  I'll modify as:

  CE1 -- PE1 -- P -- PE2 -- CE2

> 
> If the provider PE edge manually sets the precedence outbound to the CE what
> affect will this have on the TCP session.

  As I think you know, the primary catalyst for this modification
  was observed operational behaviour.  We saw that when mid-point
  devices [P] changed TOS/DSCP and the observed response packets
  [at CE1] within the TCP session varied, CE1 would reset the packet.

> the TCP RST condition should happen when their is a ["lower precedence" or
> any change in precedence] if both ends where set to the same precedence by
> the PE router 

  Although a viable solution, this is difficult to implement in
  practice.  The need this presents for PE1 and PE2 to synchronize by
  session and policy is great.  I believe this demand would severely
  impede DHCP/TOS and internet deployment.

  Another implementation workaround would maintain state as below,
  but this is only a partial resolution to the problem.

> for example precedence or dscp of "0". Does tcp security have
> a specific requirement of precedence, 

  In practice, no.  Per a strict interpretation of the RFC793, yes.
  We saw that mainly some [10%] of MacTCP (apple) TCP/IP stacks
  enforced this, as well as some old IBM stacks.

> what happens if the tcp stack
> initiates a connection with a precedence of 1 and during transmission it
> gets reassigned with precedence of "0" does this screw the TCP
> session/connection? 

  Yes, because what CE1 sends "1" is changed by {PE1,P,PE2, or CE2}
  and is not what he expected.

> is manually setting precedence or DSCP to the a single
> value outbound to the CE or TCP endpoints a possible work around?

  It might be, except that it either:

  	requires sychnornization by CE's or PE's

  	or

  	disallows tcp-based DSCP delivery.  One could set things to a
  	default value [say, 0] and remember what it got on a
  	flow-by-flow basis.  However, this would negate end-to-end
  	ability of DSCP/TCP; which is certainly a desirable behaviour
  	we do not wish to negate.

   -alan

> 
> Regards
> 	Phillip.
> 
> ********************************************
> Phillip Grasso-Nguyen (CCNA)
> Senior Network Engineer - Core Engineering Team
> Davnet Telecommunications
> Level 7, Magna Data House
> 209 Castlereagh Street, Sydney
> NSW, Australia, 2000.
> Tel: [phone] Fax: [phone]
> mailto:[e-mail]
> http://www.magna.com.au
> PGP Fingerprint:1083 7987 D33A C7E8 5DB2  AAD2 4F5D 6B99 CBB7 55A4
> PGP Key: http://www.magna.com.au/~phillipg/phillipg.asc
> Australian General Telecommunications Carrier License No 23
> ********************************************
> Disclaimer: http://www.magna.com.au/~phillipg/disclaimer.txt
> "Leave complexity at the 'edges' and keep the network 'core' simple"