Goal

1
Goal

• Providing QoS guarantee without maintaining
  per-flow state on the data or control planes.

Motivation

• Integrated Services Per-flow state at all
  routers.
• Differentiated Services Per-flow state at the
  edges and weak service semantics.
• DPS and CSFQ Per flow state at the edge routers
  and stronger service than Differentiated
  services.

Can we eliminate per flow state from the
network completely? In other words, can per-flow
state be maintained only at the endhosts?
2
Service Model

• Focus on bandwidth allocation
• Flows can reserve bandwidth (peak-rate) and then
  send data upto that rate.
• Weak service no hard delay/bandwidth guarantees

• Overview of Solution
• Protect well-behaved flows by detecting and
  containing/barring misbehaving flows.
• Data plane - recursive monitoring of aggregate
  flows.
• Control plane lightweight certificates and
  random sampling.

3
Recursive Aggregate Flow Monitoring

• Data Plane
• At a router, each flow is mapped (randomly) onto
  an aggregate flow (A-flow)aggregate flow. The
  A-Flow is monitored as a whole.
• Upon detection of misbehaving A-flow, recursive
  subdivision and monitoring of that A-flow occurs.
• In the limit, this involves policing individual
  flows of an aggregate flow.
• Resulting A-Tree size is constrained.
• Pros Low state, trade-off between state and
  time of detection.
• Cons Weaker service, time of detection, good
  only for small ratio of misbehaving flows

4
Data Plane Policing of aggregate flows

• Resources and reservations of aggregate flows
  represent the sum of resources of component
  flows.

5
Control plane

• Control plane
• Flows are admitted on a flow-by-flow basis when
  resources are available along all routers on the
  flows path.
• To enable dynamic re-mapping without per-flow
  state, is per-flow information required ?
• Aggregate flow reservations are stored as soft
  state at the routers and per-flow information is
  inferred from periodic refreshments by end-hosts.
• Lightweight certificates can be used to ensure
  that refreshes correspond to admitted flows.

6
Control Plane

• Duplicate refreshes are still a problem
• Can be checked by maintaining a small amount of
  information per flow.
• Alternative random sampling of refreshes.
• Alternative perform counting of refreshes
  received in aggregates, and compare with upper
  bounds on these counts, i.e. the counts from the
  previous time-period (These upper bounds need to
  be recalibrated periodically, by maintaining
  per-flow state within 1 or a small number of
  aggregates)