[j-nsp] A conceptual advice on QoS is needed

[Victor Sudakov]

Adam Vitkovsky wrote:
> > > >
> > > > We have Juniper switches interconnected by a transport network (NEC
> > > > equipment mostly). The nodes of the transport network are Ethernet
> > > > switches in their own right, they have 4 or 8 interface queues and
> > > > can do priority queueing and/or WRR queueing based on 802.1p
> > > > codepoints in received frames.
> > > >
> > > > Now if you had to implement QoS policies, where would you do the
> > > > actual
> > > > prioritization:
> > > >
> > > > 1. Classify the frames on the Juniper switches, but do the actual
> > > > queueing/prioritization in the transport network.
> > > >
> > > > 2. Configure one FIFO queue on the transport network nodes and do
> > > > both classification and prioritization on the switches.
> > > >
> > > > It may be important to know that most switches are connected to the
> > > > transport network by Gigabit interfaces, but the real throughtput of
> > > > the transport network itself is about 150 Mbit/s.
> > > >
> > > > But there is another branch of the transport network where all
> > > > interfaces on the nodes are 100BaseT, so the switches are connected
> > > > at
> > > > 100 Mbit/s.
> > > >
> > > > Thanks in advance for any opinions.
> > > >
> > > Victor,
> > >
> > > If your network carries traffic of higher priority that is interleaved with
> > > traffic of lower priority then the below questions are relevant:
> > >
> > > 1) Can it happen on Juniper switches that BW on input > BW on output
> > > (e.g. traffic incoming on multiple ports will be transmitted out one
> > > port)?
> >
> > Very unlikely.
> You mentioned that the Juniper switches are going to be connected to
> the transport network via 1GE interface(s) Is the sum of inputs(edge
> links) less than 1Gbps(core link) please?

Yes, it is expected to be less than 1Gbit/s, but more than 150 Mbit/s
the core can handle.

> > >
> > > 2) Can it happen on NEC transport switches that BW on input > BW on
> > > output?
> >
> > Most likely.
> >
> >
> > > If yes then you need QOS (classification and scheduling)
> >
> > The question is, where I need QoS: before the traffic enters the transport
> > switches or inside them.
> >
> If contention is most likely to happen on NEC transport switches
> then you need QOS on NEC transport switches.

It is possible to prevent congestion in the NEC transport switches by
doing shaping/policing in the Juniper switches. But should I be doing
this, that's the point of my question.

-- 
Victor Sudakov,  VAS4-RIPE, VAS47-RIPN
sip:sudakov at sibptus.tomsk.ru