A Comparison of Overlay Routing and Multihoming Route Control

1
A Comparison of Overlay Routing and Multihoming
Route Control

• Aditya Akella, Bruce Maggs, Jeffrey Pang,
  Srinivasan Seshan,
• Annes Shaikh

2
Background

• Overlay routing can improve performance over BGP,
  between a limited number of nodes RON, Detour
• Flexibility, key property of overlays
• Basic assumption of previous overlay studies is
  that end nodes subscribe to a single ISP
• Multihoming allows end-network sites to schedule
  their transfers over multiple ISPs (route
  control)
• How much benefit does overlay routing provide
  over BGP, when multihoming and route control are
  considered ?
• Round trip time and throughput metrics are used

3
Definitions

• k-multihoming End node has access to multiple
  ISPs. Emulated by a virtual node composed of
  multiple physical 1-multihoming nodes in the same
  city.
• k-overlay Same as above, but paths can use
  additional nodes
• Note Multihoming can be seen as a special case
  of overlay with paths of max 2 hops, and
  intermediate nodes on different ISPs
• Note What is the difference between 1-overlay,
  k-overlay and k-multihoming?

4
Testbed

• 68 nodes,17 US cities
• All-to-all latency and throughput measurements
• Two types of throughput estimates optimistic
  and pessimistic
• Assumptions
• Instantaneous knowledge about the performance and
  availability of routes via each of its ISPs or
  overlay paths
• For throughput maximizing overlay paths, only
  paths comprised of at most two overlay hops are
  considered

5
1-Multihoming vs. 1-Overlays

• 1-Overlays offer better round trip time (33) and
  throughput (15)
• In a large fraction of measurements (46)
  indirect paths offered better RTT performance
  than direct path.

6
1-Multihoming versus k-Multihoming and k-Overlays

• Both k-multihoming and k-overlay routing offer
  significantly better performance than
  1-multihoming
• k-overlay, (k3) performs better than 1-overlay
  (5-20 RTT)

7
k-Multihoming versus k-Overlays

• k-overlays offer marginal benefits over
  k-multihoming alone
• Only a small fraction of transfers showed
  significantly better performance with k-overlays
  rather than k-multihoming

8
Reasons for performance difference

• For large improvements (gt 50 ms) performance is
  mainly due to avoiding congestion (small
  difference in propagation delay)
• 72 of the points are above y x/2 line, which
  implies that most of the time improvement is due
  to finding physically shorter paths

9
Resilience to path failures

• All-to-all pings between 68 nodes, every minute,
  for 5 days
• A path declared lost if at least 3 consecutive
  pings are lost (3 minutes)
• Paths with more than 10 lost probes are
  eliminated from this analysis
• Epochs of failure intervals TF are computed

10
Resiliency to path failures (cont.)
11
Discussion

• Overlays vs. Multihoming (benefits and drawbacks)
• Competing or complementary ?
• Cost of operation
• Ease of deployment
• Overlay link protocols
• Systems perspective
• How to monitor different ISP connections with
  arbitrary destinations ?
• What would be a good multihoming route control
  mechanism ?
• What is a good overlay deployment strategy ?