1. QoS routing
Finding a path that can satisfy the QoS requirement of a connection.
Achieving high resource utilization.

2. QoS routing components
Maintenance of global network state
Can be done by extending OSPF to carry QoS LSAs.
Path computation
May involve multiple metrics: delay, cost, delay-jitter, packet loss rate, etc.

3. Imprecise state information?
Higher protocol overhead of QoS LSAs
More frequent updates to flood changes in resources availability.
Link up/down .vs. a 30Mbps link/a 20Mbps link.
Implication: need routing algorithms that can tolerate imprecise global network state information.

4. Maintenance of global network state information
Link state update policy:
Timer based update: LSAs are flooded periodically.
Threshold-based update with hold-down timer
LSAs are flooded when the state change passes the threshold.
Class-based update with hold-down timer
LSAs are flooded when the state change passes the class boundary.
The link state update policy can affect the characteristics of the imprecision of global network state information.
Random imprecision
Deterministic imprecision

5. Tolerate imprecise global network state information – dynamic algorithms
Safety-based routing (Apostolopoulos99)
Works for threshold and class based update policy.
compute safety (probability) based on the triggering policy.
Compute the path that can satisfy the constraint with maximum probability (safest path).
Randomized routing (Apostolopoulos99)
compute a set of feasible paths and randomly select one.
Multi-path routing (Chen98)
compute a set of feasible paths and probe all the paths.

6. Tolerate imprecise global network state information -- static algorithms
Static multi-path routing
compute a set of paths statically and probe all the paths for each request
Localized routing (Srihari00)
compute a set of paths statically and use some heuristic to determine a path to use for each request.

7. Performance Study Cycle-by-cycle network simulator
Three link state triggering policies
Five routing algorithms
Topology and traffic pattern
Various random topologies and the ISP topology.

8. Performance metric: request blocking probability.
Traffic pattern:
uniform traffic with the Poisson request arrival rate.
Constant and exponentially distributed bandwidth requirement.
Performance metric: request blocking probability.
Results are obtained with 95% confidence level.

9. Timer based policy (update interval = 5 seconds)

10. Timer based policy (update interval = 120)

11. Threshold based policy(th = 0.1, timer = 0)

12. Threshold based policy(th=0.9, timer= 0)

13. Impact of the hold-down timer(th=0.5, load=0.5)

14. Conclusion
Multi-path routing is effective in dealing with both random imprecision and deterministic imprecision.
Randomized routing performs poorly in most cases.
Static routing offers better performance than dynamic routing when global network state is extremely imprecise
Safety based routing performs well when dealing with deterministic imprecision and performs poorly when both random and deterministic imprecision is involved.