1. Coral: A Peer-to-peer Content Distribution Network
Michael J. Freedman, Eric Freudenthal, and David Mazières
In Proc. 1st USENIX/ACM Symposium on Networked Systems Design and Implementation, San Francisco, CA, March 2004.
Presented By: Devdutta Bhosale
CS 5204 Operating Systems
November 11, 2004

2. Outline What is a CDN and why do we need it ?
Shortcomings of current approach
Coral Design
Key Ideas
- DSHT, DNS, Routing and Clustering
Evaluation
Summary and Conclusion
CS 5204 Operating Systems
November 11, 2004

3. Slashdot Effect
Some sites witness a sudden surge in traffic when popular web sites posts messages/articles that contain references to them. This sudden increase in the number of hits to a web site is known as the Slashdot effect, the name derived from a well-known technology news discussion site Slashdot.
Result: Website becomes very slow or goes down and ..
Overall increase in Internet traffic.
CS 5204 Operating Systems
November 11, 2004

4. What is a Content Distribution Network
What is a Content Distribution Network ? Serve Content from a location that is ‘closer’ to a client so as to minimize load on the origin servers.
CS 5204 Operating Systems
November 11, 2004

5. Problems with CDNs
High cost of operation: $1500 for every Mbps served.
Targeted only for commercial and big websites.
Can we get the advantages of commercial CDNs for free ?
CS 5204 Operating Systems
November 11, 2004

6. Coral’s Approach
Wouldn’t it be nice to host a website capable of serving at 10 Mbps when the actual uplink capacity to the Internet is only 256 Kbps ??
Wow !! That’s great but how do we do it..
CS 5204 Operating Systems
November 11, 2004

7. Coral Approach Coral Coral Coral 9 Coral 4 8 11:Store 4 Coral Coral 72 3 5 1 10
Resolver 6
Browser
CS 5204 Operating Systems
November 11, 2004

8. Coral’s Approach
Use every node participating in the system to serve content.
Use locality of the client to serve content from a location that is closer to the client.
Avoid hot spots within Coral.
CS 5204 Operating Systems
November 11, 2004

9. Coral Components Coral consists of three main components.
Coral Indexing Infrastructure
DNSSRV
HTTPProxy
CS 5204 Operating Systems
November 11, 2004

10. Key Contributions
First Completely decentralized and self-organizing web-content distribution network.
First P2P DNS redirection
Self Organizing clusters of nodes
CS 5204 Operating Systems
November 11, 2004

11. Key Design Ideas DSHT: Distributed sloppy hash table P2P DNS System
Clustering Mechanism to exploit locality
Hotspot aware Routing
CS 5204 Operating Systems
November 11, 2004

12. DSHT Motivation With earlier DHTs one key maps to one node. i.e.
URL  Location of Node Storing the page.
If multiple nodes are storing the same URL
we can have
URL  List of nodes storing the page.
What happens when list gets updated very frequently.
CS 5204 Operating Systems
November 11, 2004

13. DSHT Motivation
Any node requesting some content is likely to cache that content which is a multiple reader-writer problem.
Allow multiple values to be associated with a key.
Each node stores only a maximum no. of values per key.
Client
Server
New Writer
CS 5204 Operating Systems
November 11, 2004

14. Coral Clusters Clusters are used to group topologically close nodes.
Coral P2P components are arranged in the form of clusters.
Each cluster is characterized by maximum desired network RTT called as diameter.
Hierarchy of different diameter clusters is known as levels.
Every node is a member of one cluster at each level.
There are multiple clusters of levels 2 and 1while there is only cluster of level-0.
Cluster arrangement changes dynamically. (split, merge)
CS 5204 Operating Systems
November 11, 2004

15. Coral Clusters
CS 5204 Operating Systems
November 11, 2004

16. Coral Clusters
CS 5204 Operating Systems
November 11, 2004

17. P2P DNS Server probes client
Server probes client to determine the RTT and then return a proxy who is below that level.
Server probes client
Browser
Coral
httpprx
dnssrv
Resolver
Coral
Proxy
CS 5204 Operating Systems
November 11, 2004

18. P2P DNS
The closer the dnssrv to client the better is the selection of CoralProxy.
In the AUTHORITY section of the DNS reply it sends the DNS server addresses which are close to the client so that future DNS requests will be resolved by that server.
CS 5204 Operating Systems
November 11, 2004

19. Coral Routing - XOR Metric Nodes and Keys are both 160-bit.
Distance between a key and a node is XOR of their IDs.
IDs with longer matching prefixes are numerically closer.
If a distance has one fewer bit than other it means it is at least halfway close. (Binary Property)
CS 5204 Operating Systems
November 11, 2004

20. Coral Routing 1110 XOR 0100: 1010 0000 XOR 0100: 0100
NODEID KEY: Distance
1110 XOR 0100: 1010
0000 XOR 0100: 0100
0101 XOR 0100: 0001
0100 XOR 0100: 0000
CS 5204 Operating Systems
November 11, 2004

21. How does Coral Prevent Hotpots ?
Store value once in each level cluster
Always storing at closest node causes hotspot
Halt put routing at full and loaded node (use of stack)
Full → M vals/key with TTL > ½ insertion TTL
Loaded → β requests in past minute
Cache Key/Value pairs at a node close to the key being referenced.
CS 5204 Operating Systems
November 11, 2004

22. Evaluation
We can evaluate Coral by how it achieves following objectives
1. Reduction in load on servers, solving the “flash crowd” problem.
2. Clustering provides performance gains for popular
data, resulting in good client performance.
3. Ability to naturally forms suitable clusters.
4. Preventing hot spots within its indexing system.
CS 5204 Operating Systems
November 11, 2004

23. Experimental Setup
On 166 machine in North America and Europe, Coral Daemons, DNS servers and HTTP proxies are started.
Origin server has an upstream bandwidth of 384 kbps serving 12 different 41 kb files.
Client processes are launched on all the machines requesting images for 30 minutes.
Total load on the system is Kb/s.
CS 5204 Operating Systems
November 11, 2004

24. Server Load Coral Hits in 20ms Cluster Local Caches begin to
Handle requests.
Hits to Ori-
-gin server
CS 5204 Operating Systems
November 11, 2004

25. Summary and Conclusions
Provides non-standard P2P storage abstraction
Prevents hotspots
Solves Slashdot effect
Cheap option to traditional CDNs
CS 5204 Operating Systems
November 11, 2004

26. Cons
Difficult to guarantee a minimum quality of service as offered by tradition CDNs
Cannot serve dynamic content
Security
Behavior of the system varies as the nodes join and leave the system
Latency: 50 % clients experienced a latency of 3 seconds but paper does not talk about the latency without Coral
CS 5204 Operating Systems
November 11, 2004

27. Questions ??
CS 5204 Operating Systems
November 11, 2004