EE 122: Lecture 22 (Overlay Networks)

1
EE 122 Lecture 22(Overlay Networks)

• Ion Stoica
• November 27, 2001

2
Motivations

• Changes in the network happen very slowly
• Why?
• Internet network is a shared infrastructure need
  to achieve consensus (IETF)
• Many of proposals require to change a large
  number of routers (e.g., IP Multicast, QoS)
  otherwise end-users wont benefit
• Proposed changes that havent happened yet on
  large scale
• Congestion (RED 93) More Addresses (IPv6 91)
• Security (IPSEC 93) Multi-point (IP multicast
  90)

3
Motivations (contd)

• One size does not fit all
• Applications need different levels of
• Reliability
• Performance (latency)
• Security
• Access control (e.g., who is allowed to join a
  multicast group)

4
Goals

• Make it easy to deploy new functionalities in the
  network ? accelerate the pace of innovation
• Allow users to customize their service

5
Solution

• Deploy processing in the network
• Have packets processed as they traverse the
  network

IP
Overlay Network (over IP)
AS-1
6
Examples

• Overlay multicast
• Increase robustness and performance
• Content Distribution Networks (CDNs)

7
Motivations IP Multicast Problems

• Scalability with number of groups
• Routers need to maintain per-group state
• Aggregation of multicast addresses is complicated
• Supporting higher level functionality is
  difficult
• IP Multicast best-effort multi-point delivery
  service
• Reliability and congestion control for IP
  Multicast complicated
• Need to deal with heterogeneous receiver ?
  negotiation hard

8
Approach

• Provide IP multicast functionality above the IP
  layer ? application level multicast
• Challenge do this efficiently
• Projects
• Narada
• Overcast
• Scattercast
• Yoid

9
Narada Yang-hua et al, 2000

• Multi-source multicast
• Involves only end hosts
• Small group sizes lt hundreds of nodes
• Typical application chat

10
Narada (contd)
Stanford
Gatech
Stan1
Stan2
CMU
Berk1
Berk2
Berkeley
Overlay Tree
Stan1
Gatech

Stan2
CMU
Berk1
Berk2
11
Discussion

• Scalability ( of groups)
• Routers do not maintain per-group state
• End systems do, but they participate in very few
  groups
• Easier to deploy
• Potentially simplifies support for higher level
  functionality
• Leverage computation and storage of end systems
• For example, for buffering packets, transcoding,
  ACK aggregation
• Leverage solutions for unicast congestion control
  and reliability
• Scalability ( of receivers) still an open issue
• Other solutions (e.g., Overcast) are scalable but
  not as flexible typically assume single-source
  multicast trees

12
Examples

• Overlay multicast
• Increase robustness and performance
• Content Distribution Networks (CDNs)

13
Motivation

• Routing in the Internet is not optimal with
  respect to
• Performance packets do not necessary propagate
  along the shortest path
• Robustness two nodes may not be able to
  communicate although there is a path between them
• Why?

14
Solution

• Control routing at the application level
• Projects
• Resilient Overlay Networks
• Detour

15
Resilient Overlay Networks Anderson et al, 2001

• Make the end to end communication more robust
• Each node monitor the network conditions to every
  other node by periodically probing the network
• If node n1 cannot reach n2 directly, try to reach
  it through an intermediate node n3
• Intended application robust communication in a
  small group (lt 50, 60 nodes)

16
Resilient Overlay Networks (contd)

• N1 can no longer communicate directly to N2

N2
N1
17
Resilient Overlay Networks (contd)

• Find a node N3 such that N1 can communicate with
  N3 and N3 with N2

N2
N1
18
Discussion

• Find an alternate path in most cases when two
  nodes cannot communicate directly
• Can be used to provide better delay and bandwidth
  than the direct IP route between two nodes
• Scalability still an open issue

19
Examples

• Overlay multicast
• Increase robustness and performance
• Content Distribution Networks (CDNs)

20
Motivations

• Todays Internet is not optimized for Web traffic
• Many clients transfer the same information (e.g.,
  CNN front page, software downloads)
• Identical files are transferred over and over
  again
• IP multicast not a solution
• Users dont access the same info at the same time
• Users have widely different capabilities
• Communication cable modem vs. dial up modem
• Display high-resolution workstation monitor vs.
  Palm Pilot

21
Solution

• Have nodes inside the network that store and
  process the documents
• Examples web caching, transcoding

22
Base-line Solution

• Many clients transfer same information ?
• Generate unnecessary server and network load
• Clients experience unnecessary latency

Server
Backbone ISP
ISP-1
ISP-2
Clients
23
Reverse Caches

• Cache documents close to server ? decrease server
  load
• Typically done by content providers

Server
Reverse caches
Backbone ISP
ISP-1
ISP-2
Clients
24
Forward Proxies

• Cache documents close to clients ? reduce
  network traffic and decrease latency
• Typically done by ISPs or corporate LANs

Server
Reverse caches
Backbone ISP
ISP-1
ISP-2
Forward caches
Clients
25
Content Distribution Networks (CDNs)

• Integrate forward and reverse caching
  functionalities into one overlay network
  (usually) administrated by one entity
• Example Akamai
• Documents are cached both
• As a result of clients requests (pull)
• Pushed in the expectation of a high access rate
• Beside caching do processing, e.g.,
• Handle dynamic web pages
• Transcoding

26
CDNs (contd)
Server
CDN
Backbone ISP
ISP-1
ISP-2
Forward caches
Clients
27
Discussion

• CDNs were developed to efficiently handle todays
  web traffic
• Relive server and network load
• Perform load balancing, caching
• Increase client performance
• Process data according to clients needs
• A basic technique that makes CDNs possible is
  redirection (see HTTP). How?

28
Conclusions

• Overlay networks allow to deploy new services in
  the network today, e.g.,
• Multicast, CDNs
• Can increase network robustness and client
  perceived performance, e.g.,
• RON, CDNs
• Challenges
• Efficiency
• A packet may need to be processed above transport
  layer before reaching the destination
• Path followed by the packet might be worse than
  the direct IP route
• Scalability