Content Distribution Network Coral system NSDI 2004

1
Content Distribution Network(Coral system NSDI
2004)
2
Background

• DHT
• Key space
• Overlay network
• Web caching
• proxies
• Content Distribution Network
• e.g. Akamai
• Problem being addressed by Coral
• Scalable content distribution
• Avoid slashdot effect

3
Takeaway

• Novel key/value indexing called Coral
• Weaker consistency than traditional DHT
• DSHT where S stands for Sloppy
• Locality
• Hot spot prevention
• CoralCDN
• Self-organizing decentralized
• Locate nearby cached copies without querying
  distant nodes
• Avoids slashdot effect
• Interoperates with unmodified web browsers

4
Coral Indexing DSHT

• Nodes are organized into clusters
• Cluster
• Maximum desired RTT (diameter)
• Hierarchy of diameters levels
• Level 0 infinity
• Level 1 60 ms
• Level 2 20 ms
• API
• Put (key, value, ttl, levels)
• Get (key, levels)
• Nodes (level, count, target, services)
• Levels()

5
Coral System
6
CoralCDN components

• DNS server
• Returns Coral proxies for client query from a
  browser (through a resolver)
• Coral HTTP proxy
• Proxy selection based on client locality
• Get from peer proxies (using Get) if at all
  possible
• Goal is to reduce load on origin server
• Flash crowd
• Automatically form a (sort of) multicast tree

7
Coral key based routing
8

• Basic idea in routing
• Reduce distance to desired destination by
  (approx) half in each iteration
• Reduces overloading nodes close to the key
• Potential longer latency during lookup
• Tradeoff to avoid tree saturation (due to
  slashdot effect)

9
Sloppy storage

• Put and Get in other DHT
• All proceed to node closest to key
• Coral
• Often satisfied by nodes other than those close
  to key
• Two phase put
• Node attribute full, loaded
• Forward phase
• Putter send RPCs to nodes en route close to key
• Stop at full and loaded node
• Reverse phase
• Pick a node that is closest to the key as the
  destination for the put
• Multiple values may be stored for a key

10

• Get (retrieval)
• Remote peer (for k) returns values
• Application dependent action on values returned
• E.g. Coralproxy
• Values are locations of files that hash to k
• parallel download of files from the multiple
  locations

11
Hierarchical operations

• Hierarchical retrieval using clusters

12

• Hierarchical insertion
• Start put in level 2 cluster
• Continue putting in level 1 cluster
• Loose hierarchical cache

13
Cluster management

• Joining node contacts an existing peer node to
  join
• A node will join only an acceptable cluster
• Latency limit for the cluster and RTT results to
  the members of the cluster
• All RPC results cluster contain meta data for
  fine tuning the cluster membership

14
Implementation

• Coral
• 14K lines of C
• DNSserv
• 2K lines of C
• HTTP proxy
• 4K lines of C

15
Evaluation

• Load distribution and control of flash crowd
  problem
• Clustering
• Ability to self-manage clusters
• No hot spots in the indexing system

16
Experimental setup

• 166 Plantelab machines seeded with a host name
• Run coral daemon, dnsserv, coralproxy on each
• Unmodified Apache web server on host
• Clients on the planetlab machines start
  generating requests

17
Server load results
18
Client latency
19
Clustering