Replica Placement Heuristics of Application-level Multicast

1
Replica Placement Heuristics of Application-level
Multicast

• Chia-Hsing Yu
• Jiahua He
• CSE of UCSD

2
Outline

• Multicast and RMX
• Model and Heuristics
• Simulation and Results
• Conclusion and Future Work

3
Application-level Multicast

• Goal
• Distribute Contents to Many Clients
• Problem
• How to reduce the load of the central server?
• How to reduce the response time of requests?
• Replication at different servers

4
RMX Reliable Multicast proXy
TCP
SRM Reliable IP Multicast
5
RMX

• Semantic reliability
• information ?? representation of information
• Sender can lower the stream resolution if the
  network load is heavy

6
Existing Problems

• Only sources, no replicas
• No request, only recovery request
• Static RMXs in network
• Static configuration of data groups

7
Related works

• Replication in unstructured P2P (Princeton)
• Owner, Path, Random
• PAST(Microsoft and Rice)
• Nodes with similar ids
• OceanStore (Berkeley)
• On or near the clients
• Focus on persistent storage with versions
• Chain (Cornell)
• Machines with replicas of a same file form a
  chain
• Focus on availability

8
Model and Heuristics

• Fixed sources and dynamic replicas
• Streaming multicast on demand
• No replication
• Baseline
• Replication on path
• FIFO
• LRU
• Color

9
Baseline

• Only sources, no replicas
• Learning bridge scheme to search
• Learn routing information from incoming data
• Soft state periodically refresh
• Request suppression
• Ideal condition no loss

10
FIFO and LRU

• Replication on path
• Broadcast to search
• FIFO
• Remove the oldest one if no space
• LRU
• Order the files by last usage
• Remove the oldest one if no space

11
Color

• Graph coloring
• Neighbors with different colors (files) from mine
• Can get more different files from neighbors
• Remove the file with nearest replica
• Visiting Frequency
• More frequently visited, more possible to be
  visited
• Cost function dist freq
• dist distance to the nearest replica
• freq visiting frequency
• Upper bound of the cost if removed

12
Simulator

• Event-driven Simulator

New Event
New Event
New Event
Event Handler
Min Heap
Earliest Event
13
Simulator(2)

• Stream-level Simulation
• SIM_SEND_STREAM( bit rate, length )
• Input
• Network Topology
• Host Resources
• Stream Sources
• User Requests

14
Experiment Configuration

• Network Topology
• Binary Tree
• Host Resources
• 127 hosts (data groups)
• Hard disk size variable
• Stream Sources
• 1270 sources (average 10 per host)
• 500 Kbps, 8000 seconds each
• Randomly distributed
• User Request
• Randomly distributed
• Total number variable
• Experiment Span
• 100 hours

15
Experiment Configuration (2)

• Variances
• Number of requests 211 218
• Hard disk size 8G 128G
• Metrics
• Client view average response time
• Server view load (number of streams per
  RMX) load balance (standard deviation of load)
• System view throughput

16
Client View Avg. Response Time vs. of Requests
About 30 improvement
17
Client View Avg. Response Time vs. Disk Size
Disk size outperforms replication strategy
18
Server View Avg. of Streams vs.
of Requests
About 50 improvement
19
Server View S.D. of Streams vs.
of Requests
About 50 improvement
20
Server View Avg. of Streams vs.
Disk Size
Disk size outperforms replication strategy
21
Server View S.D. of Streams vs.
Disk Size
Disk size outperforms replication strategy
22
System View Throughput vs.
Requests
About 25 improvement
23
System View Throughput vs. Disk
Size
Upper bound 25.4398
24
Contributions

• Implement and analyze Baseline, FIFO, LRU algs
• Propose and verify Color heuristics
• Avg. response time up to 30 improvement
• Load up to 50 improvement
• Load balance up to 50 improvement
• Throughput up to 25 improvement

25
Future Works

• Biased requests
• Heterogeneous environment (hosts, links, streams)
• Random forward
• More sophisticated heuristics
• Experiment in real environment