Walking the Tightrope: Responsive Yet Stable Traffic Engineering

1
Walking the Tightrope Responsive Yet Stable
Traffic Engineering

• Presentation by Umamaheswararao K

2
Traffic Engineering methods

• Goal Minimize the max utilization in network
• Offline Methods
• - OSPF-TE, Oracle
• Online Methods
• - MATE, TeXCP (current approach)
• -Common Practice InvCap

3
Offline Methods

• Use long term average traffic demands to route
  traffic
• Cannot react to real time traffic changes
• May not prevent congestion when unanticipated
  combination of failures occur
• suboptimal/inadequate load distribution
• Eg OSPF-TE, OSPF-TEfailure, Oracle

4
Online Methods

• Balances load in real time
• Responds to changes in TM and failures
• Eg TeXCP, MATE
• Challenges
• Stability and performance
• Quick recovery
• Low network cost
• Practical to implement

5
TeXCP
6
TeXCP
7
Variables Used
8
TeXCP

• Distributed, responsive, Stable Online Protocol
• Two components
• Load Balancer shifts traffic from one path
  to another based on network state
• Closed Loop feedback controller collects
  network feedback for stability

9
How does it work?

• TeXCP agent at each Ingress router for each IE
  flow
• - Uses light-weight feedback from core routers
  about path utilization
• - Traffic split at a granularity of flow

10
How do we evaluate?

• For same TM, utilization of network and failure
  resilence
• How far is it from optimal utilization value
• Compare with MATE
• Pruning of paths that donot affect max util

11
Problem formalization
12
TeXCP agent

• Ingress router runs an agent for each IE pair
• Probes each path for link util and failure
• Splits the traffic across paths based on util
• Adapts split ratios in real time
• All agents are independent , no xchange of info
• Independent of routes

13
Path selection

• ISP configures set of paths for each agent
• Picks K-shortest paths for IE pair
• Paths are computed offline (based on topology)
• Agent uses any path based on its state

14
Probing network state

• Sends a probe across all paths every probe timer
  period Tp gt RTT
• Default Tp 100ms
• Each router updates the link util in the pak as
  Max(its util, util so far seen in the path)
• Processed at slow path at each router
• How to handle loss?

15
Load balancer

• Recalculates load distribution every decision
  timer Td gt 5Tp

16
Stability (preventing Oscillations)
17
Solution

• Computing aggregate feedback
• - done at every Tp
• Computing per-IE-flow feedback
• Sending the feedback to TeXCP agents
• Shorter paths are prefered

18
Various TE techniques
19
Results and comparision
20
Under failures
21
Comparison with MATE
22
(No Transcript)
23
Diff with MATE

• Texcp minimizes max util
• MATE minimizes sum of delays
• Texcp distributed, MATE assumes ingress node has
  instantaneous knowledge of nw
• MATE doesnot require core routers to report util

24
Convergence Time

• Topology, max path per IE flow, split ratio diffs

25
Related work

• Optimal Routing
• Offline Oblivious Optimizers
• Online Load balancers

26
Adv/Disadv

• Adv many we discussed
• Disadv Can increase the delay
  Reordering of TCP paks
• Large number of control paks
• Scalability (Timers Tp and Td, agents)
• Flow state at edge router
• Processing overhead at each router

27
Future work

• Interaction between overlay and online TE
• Reducing flow state
• Addressing scalability issues with timers/agents
• Application based routing
• Reducing convergence time