Dynamics of End-host controlled Routing

1
Dynamics of End-host controlled Routing

• Mukund Seshadri
• Prof. Randy Katz

Sahara Retreat Jan 2004
2
Problem

• Consider multiple independent overlay
  networks/flows, each choosing the best overlay
  route
• Can this be unstable/inefficient?
• Identify such scenarios.
• Suggest improvements.

3
Motivation

• Overlay routing can provide better functionality,
  performance or resilience.
• e.g. RON3, Detour4, ESM5.
• What if several entities set up their own overlay
  flows?
• e.g. using overlay support primitives 2.
• Primary app multimedia streams.
• Flows can have some physical links in common, no
  explicit coordination.
• e.g. on popular shared test-beds like
  PlanetLab6.
• Different networks/independent flows from same
  network.

4
More Background
Extra Slide

• Resilient Overlay Networks
• Recovers from routing failures in around 20s, as
  opposed to several minutes in normal BGP.
• If default route from Node A to B fails, then
  data is redirected through Node C.
• All available paths are probed frequently
• Does not scale beyond 50 nodes
• End System Multicast
• End-hosts form a low-delay or low-b/w
  degree-bounded mesh and then a multicast tree.

5
Unstable Routing Example
Data Paths
Available Paths
Bottleneck Physical Link
Overlay Nodes
Each source has a 1Mbps flow.

• Oscillation of both flows between the two
  alternate paths is possible.

6
Outline of Study

• Used simulations to study requirements for good
  performance and factors affecting it.
• Some form of restraint is needed
• Hysteresis Threshold (H)
• Randomized selection
• Decision times.
• Automatic discovery of H
• Factors affecting performance
• Size and number of flows, path density,
  cross-traffic, more

7
Simulation Model

• M overlay networks/flows with N available overlay
  paths each
• All paths monitored
• Available b/w inferred perfectly in a time
  window (Tm)
• Configurable error factor
• Best path is selected to send traffic on (GREEDY)
• Route change based on bandwidth improvement
  threshold (H)
• Periodic decisions (Tr)
• M 100-1000, N 5-50.
• Path-level simulator
• Characterizes shared bottleneck links.
• The level of sharing is characterized by path
  density (Pf)
• Unicast CBR flows with bandwidth requirement.
• Flows arrive and depart with lifetimes around
  1000 sec.
• Metric Loss Rate (related to bandwidth).

8
Simulation Parameters
Extra Slide

• Unless mentioned otherwise, these are the values
  used for system parameters.

9
Need for Hysteresis

• No/Low hysteresis gt very unstable, high loss
  red line.
• H too high gt high loss due to poor route
  selection blue line.
• Optimal value of H green line.

10
Path density and H

• The best value of H varies significantly with
  path density (Pf), flow size and other
  parameters.

• The minimum of each each line is the best setting
  of H, for that value of Pf .
• High Pf gt greater chance of interaction gt worse
  stability and loss rate.

11
Other factors and H
Extra Slide

• The best value of H varies with other parameters
  too
• Relative flow size proportional to
  inter-arrival-time (IAT).
• Cross-traffic percentage.
• Explanation of observed trends the impact of a
  flows re-routing is more significant when a it
  is a larger fraction of link capacity.

12
Other factors - Summary

• Combination of following factors leads to poor
  performance
• High path density
• Large flow size and number
• Low cross-traffic
• High load
• High variation in bandwidths.

13
Routing window (Tr) and H

• Increasing routing window while keeping
  measurement window constant can improve
  performance
• Since the no. of flows re-routing during the
  measurement window decreases.
• But can increase reaction time after failure.
• Best value of H (line minimum) varies a lot

14
Improvements to GREEDY

• Randomly select path to be chosen
• ARAND In proportion to available bandwidths
• SRAND Best of randomly selected subset of size S
• in proportion to capacity
• Reduces measurement overhead
• Works well for server load balancing 1
• (different work model jobs are assigned to only
  one server for their lifetime)
• GRAND Randomly select from the best S paths

15
Randomized Selection

• Much lower loss than GREEDY
• SRAND and ARAND best

• Best value of H still varies w.r.t. path density,
  etc, for SRAND

16
Automatic Discovery of H

• We propose that flows automatically discover the
  most suitable values of H.
• Flows can independently probe values of H
• No route change gt decrease H
• Route change gt increase H
• MIMD works slightly better than other methods.
• High initial value quick-start decrease phase
  (high decrease factor)

17
Performance of H-discovery

• Very low loss-rates compared to fixed-H.

• Upper edge of C.I. is much lower than GREEDY.

• H-discovery works well in all scenarios,
  including high IAT, below.

18
Cheating flows

• What if most flows use restrained selection,
  while some cheat by using more aggressive
  GREEDY methods?
• When good flows use fixed H
• The cheaters obtain much lower loss rates
• Good flows dont suffer unless cheaters exceed
  35 of all flows.
• When good flows use H-discovery
• The cheaters do not benefit
• Good flows loss increases when cheaters exceed
  20 of all flows, but the loss is still lower
  than with fixed-H.

19
Cheating flows - Graphs
Extra Slide

• The cheaters benefit when good flows use fixed
  H
• but not with the H-discovery method

20
Conclusion

• Already summarized the effect of different
  factors on performance.
• Restraint is useful in route selection.
• H, randomization, Tr
• We propose dynamic discovery of H
• Low loss rate in all scenarios.
• Future work
• Investigate dynamic models of flow and
  cross-traffic.
• Study the usefulness of these forms restraint in
  network-layer routing.

21
References
Extra Slide

1. How Useful is Old Information M.Mitzenmacher
   PODC 1997
2. Infrastructure Primitives for Overlay Networks
   Karthik Lakshminarayanan et al. under
   submission.
3. Resilient Overlay Networks Andersen et al
   SOSP 2001
4. Detour a Case for Informed Routing and Transport
   Savage et al. IEEE Micro Jan 1999.
5. A Case for End System Multicast Yang-hua Chu et
   al. JSAC 2002.
6. PlanetLab http//www.planet-lab.org