[Irtf-rr] AW: Differentiated Routing, not only plain rambo-SPF
Hummel Heinrich
Heinrich.Hummel@icn.siemens.de
Mon, 12 Aug 2002 17:02:57 +0200
Jing,
Let me recapture and use the following DSCPs:=20
DSCP-Gold-1st, DSCP-Gold-2nd, =
DSCP-Silver-1st,DSCP-Silver-2nd,DSCP-Bronze-1st,DSCP-Bronze-2nd.=20
DSCP-Gold-1st and DSCP-Gold-2nd shall trigger DiffServ Gold-SLA =
handling.
DSCP-Silver-1st and DSCP-Silver-2nd shall trigger DiffServ Silver-SLA =
handling.
DSCP-Bronze-1st and DSCP-Bronze-2nd shall trigger DiffServ Bronze-SLA =
handling.=20
They shall as well identify road systems:=20
DSCP-Gold-1st, DSCP-Gold-2nd, =
DSCP-Silver-1st,DSCP-Silver-2nd,DSCP-Bronze-1st,DSCP-Bronze-2nd - built =
in this
sequence. Hereby each road system is based on what has been left over =
by the previously built road systems (in terms of
bandwidth).
Gold: A smallest size tree of toll roads, highways, state roads, =
country roads.
Silver:A smallest size tree of highways, state roads, country roads.
Bronze:A smallest size tree of state roads, country roads, and highways =
only if no path otherwise.
=20
QoS parameters (speed, packet loss rate, MTU size,..) may define the =
type of road.
1st and 2nd road system may be different at least w.r.t. their kernel =
parts. (The kernel part be that=20
center of the smallest size tree, which is used by at least m =
ingress/egress combinations).
Either traffic starts out with a) proper DSCP-Y-1st, resp. DSCP-Y-2nd, =
or b) is changed from DSCP-Y-1st to
DSCP-Y-2nd at the rim of the kernel if some current traffic load =
information advises so. I prefer a) and if possible such
that any microflow will entirely take one single route.
Jing, you are mentioning two aspects, namely stability and =
extensibility.
Stability:=20
We may toss a dice at the ingress,i.e. determine a random number Y =
between 0.0 and 1.0. If =20
0.0<=3D Y < a, then we use DSCP-Y-1st, if a<=3D Y < 1, then we use =
DSCP-Y-2nd. The real number a should be such
that we avoid/dissolve some congestion and do not relocate it. At =
low-traffic times a is equal 1. But for the busy hour
determining value a will be a very interesting topic.=20
If any University folks were interested in doing simulations as to =
determine value a for all ingress/egress combinations
and for all three SLAs, I would be happy to exchange more thoughts and =
details on it (so far I have always deferred it).
Especially, it would be interesting to get comparable results between =
what I called "Orchestrally Conducted Traffic" and
OMP.
However we would be leaving the DiffRout-track of thoughts. Therefore =
we should say: this is out of scope of this thread.=20
It may only be addressed by some other thread.
Extensibility:
So far I have no solution, on how to determine a smallest size tree =
made out of OSPF-links and EBGP/IBGP-hops.
BTW:
It seems to be appropriate to focus on smallest size tree road =
systems, at least for some time, although ring and
somehow meshy road systems might also be considered.=20
Can anybody contribute some solution ? I think, it must be an =
identical tree of BGP-hops. Only all edge routers of the
same OSPF network must aquire the completely identical tree. Somehow:-)
Heinrich=20
-----Urspr=A8=B9ngliche Nachricht-----
Von: shen jing [mailto:jshen@cad.zju.edu.cn]
Gesendet: Dienstag, 6. August 2002 15:04
An: Hummel Heinrich
Cc: irtf-rr@puck.nether.net; routing-discussion@ietf.org
Betreff: Re: Differentiated Routing, not only plain rambo-SPF
Hi Hummel,
As DSCP is used for packet queueing in routers, there may be many =
types of queueing mechanism for scheduling of
different classes.
As mentioned in previous message that when overflow/congestion happens =
along some path, the flow of the same DSCP
could be overflowed onto other path (?) or it could be dispersed onto =
multiple paths computed.
This means, if we want to adapt to time-varying network status LSA =
must be triggered by overflow of any queue on its
outgoing interface. Because internet traffic's self-similarity =
characteristics and "heavy-tailed" flow sized distribution,
we
could not expect the network status maintains a stable status. That is, =
routing behavior have to be adjusted frequently.
Right ? CBQ or the like may adjust the resource allocated dynamically =
according to load on routers, how could that be
reflected ? I think, all these factors will do harm to routing =
stability and extensibility, esp. when we consider
inter-AS routing.
Well, I think some time to read those paper suggested.
>
>
> Hummel, was your original proposal same like Diffserv-aware-TE. I =
thought
> it was different.
>
> My answer:
> DiffRout is definitively different.
>
> My motivation and my key thoughts to open this discussions have been:
> - combine the best of Intserv/MPLS and Diffserv and as a result do =
stateless IP forwarding but get
> what Intserv/MPLS are providing: Qos/BW/Policy/Traffic =
balancing/rerouting.
>
> - therefore construct road systems and do IP forwarding/routing also =
based on DSCP.
>
> - the road systems may have assigned bandwidth. This is =
reasonable/feasable because some traffic can really
> be locked into some slim road system (smallest size tree, smallest =
ring) just like MPLS traffic can be locked into
> some LSP. Whereas with rambo-SPF bandwidth assignment makes no =
sense as the sum of all
> shortest path trees computed by all the nodes would span the entire =
network, i.e. spread traffic all over it.
> The price for staying within some road system should be acceptable: =
Eventually, make a few hops more.
> Based on my computations, the smallest size tree may be smaller =
than the avarage shortest path tree by 15-20 %.
> So I think the detour for the packets will not be very big either.
>
> - be guided by what you can observe on daily road traffic: Especially =
in the U.S. there are so many ramp-based
> road junctions, which are more impressive than any traffic light =
controlled junction, so that motor-cyclists may
> jump the queue of waiting cars (=3DDiffServ) if the traffic light =
is red.
> There are so many observations that can be made watching how the =
traffic is managed for cars,
> trains, airplanes, pedestrians, etc. We all are used to it and =
don't discern extra smartness in it. However there is.
>
> - and I realized what is the big obstacle: Yes it is this rambo-SPF =
routing paradigma. It does not represent the
> surveillant view, only the egomanic view, combined with the hope =
that in total you get best results because
> "the streets are emptiest when each car takes the shortest =
route":-).
>
> Heinrich
>
>
>
>
>
> _______________________________________________
> routing-discussion mailing list
> routing-discussion@ietf.org
> https://www1.ietf.org/mailman/listinfo/routing-discussion
--
Jing Shen
State Key Lab of CAD&CG
ZheJiang University(YuQuan)
HangZhou, Z.J. 310027
P.R.China
Tel: +86-571-87932423
Email: jshen@cad.zju.edu.cn
**********************************************************************
* The SunShine of life is made up of very little beams which is *
* bright all the time *
**********************************************************************
_______________________________________________
routing-discussion mailing list
routing-discussion@ietf.org
https://www1.ietf.org/mailman/listinfo/routing-discussion