NeXtworking

1
Traffic modelling and traffic control ...

• ... are necessary for QoS (or qos) and this is
  still an issue
• (over)provisioning...
• ... and controlled sharing
• QoS tends to mean "hard" qos (Intserv-like)
• for an announced connection a priori traffic
  descriptor, admission control, scheduling and/or
  traffic conditioning
• statistical network calculus is the state of the
  art in realizing hard QoS guarantees (delay lt d,
  negligible loss)
• Diffserv is softer qos
• aggregate descriptors, conditioning and per hop
  behaviours
• but how can we meet the SLA's?
• some believe even softer qos is sufficient (and
  necessary?)
• a self-managed Internet...
• ... or a network with implicit admission control
  and implicit differentiation
• we also need modelling to understand best effort
  control!
• eg, TCP modelling and re-design

2
Answering the questions

• is standard queueing theory enough?
• frequently model at the appropriate granularity,
  seek insensitivities
• is first course control theory enough?
• significant advances cf. work on stability of
  TCP
• how can we understand time-scale interactions and
  design for them?
• time scale separation is key to successful
  modelling
• but beware of drawing conclusions without regard
  to all time scales
• are time varying averages enough, or do we need
  to capture higher moments?
• insensitivity results suggest time averages are
  often sufficient
• stationarity is a key assumption is this valid
  with LRD traffic?

3
Answering the questions

• what happens to "effective" bandwidth and good
  old Erlang?
• effective bandwidth is useful for sizing
• Erlang is still insensitive ... to criticism!
• why are we insisting on modelling TCP and RED
• evolution to higher rates, ensure stability,
  improve fairness? improve efficiency?
• why did it take us some time to be more careful?
• self-similarity vs pseudo ss, limit cycle vs
  instability, etc
• who was not careful? who is now more careful?