The MPLS WG Archive

[Jay Wang <jawang@cosinecom.com>]

Title: RE: QoS and Labels: A question ??

Jim,

> -----Original Message-----
> From: Jim Boyle [mailto:jboyle@Level3.net]
> Sent: Wednesday, January 24, 2001 8:13 AM
> To: David Charlap
> Cc: mpls@UU.NET
> Subject: Re: QoS and Labels: A question ??
>
>
>
>
> That is the difference between diffserv and intserv, I
> believe. Diffserv doesn't aim to give hard guarentees to individual flows. 
> Diffserv, and "E-LSPs", work.
>

I would not compare Intserv and Diffserv like that. I think the idea of Diffserv
is to provides the needed transport, on an aggregated basis (for better scalability
of course), to achieve services that are specified by Intserv model (not exclusively
of course). Even the Diffserv paradigm is built upon PHB and aggregation concepts,
it is still possible (and very feasible) to achieve per-flow QoS guarantee.

- Jay

> Jim
>
>
> On Wed, 24 Jan 2001, David Charlap wrote:
>
> > Bora Akyol wrote:
> > >
> > > >>>>> "David" == David Charlap <david.charlap@marconi.com> writes:
> > >
> > >     David> If you need actual QoS guarantees, or if your
> > >     David> switches only support L-LPSs, or if you want the two
> > >     David> service levels to be routed independantly, then you
> > >     David> need to create a separate LSP for each service level.
> > >     David> CR-LDP or RSVP-TE can be used for signalling the QoS
> > >     David> requirements for the two LSPs.
> > >
> > > David
> > >
> > > You are making assumptions about a particular switch
> > > architecture when you assert that you can not do "actual" QoS
> > > guarantees with E-LSPs. If you have a policer and a
> > > shaper/scheduler that can process the EXP bits as well as the
> > > label, then doing "actual" QoS is quite possible with
> > > E-LSPs.
> > >
> > > Not only that, but you can map traffic destined to the same
> > > prefix to different EXP levels based either on Diffserv bits or
> > > on packet content. This also applies to VLAN tags.
> >
> > The meaning of the EXP bits is not signalled.  It must be
> preconfigured
> > network-wide.  And in order to be useful, they must be
> identical across
> > the entire network.  Which limits you to 8 different service levels,
> > network-wide.
> >
> > If you use the EXP bits to define a priority level and two flows are
> > both assigned that level, they will be queued in best-effort fashion
> > relative to each other.  There is no way to guarantee the individual
> > flows any particular level of service.  You can only guarantee the
> > aggregate of one class relative to another class.
> >
> > Using L-LSPs, you can request a separate label, and a
> separate resource
> > reservation for every flow, if needed.  This is the only way to
> > guarantee this level of service.  (Unless you're willing to
> limit each
> > DS code point to a single flow, which is a ridiculous concept.)
> >
> > E-LSPs are useful, no doubt about that.  But they can not
> provide the
> > level of QoS necessary to make sure that a single specific
> flow can be
> > guaranteed the resources it requires.
> >
> > -- David
> >
>