Cooperative Multimedia Proxy Servers

1
Cooperative Multimedia Proxy Servers

• Dr.Philip Tse
• The University of Hong Kong

2
Introduction

• Taxonomy
• Our design
• Simulation results
• Summary and Work to continue

3
Cooperative Multimedia Proxy Servers
B
A
objects
C
Server
Storage System
Proxy servers A, B, C pass their cached objects
to each other
4
Taxonomy of Cooperative Multimedia Proxy Caching

• Object partitioning and transformation
• Proxy cooperation mechanisms
• Proxy cache admission and replacement policies

5
Object Transformation and Partitioning

• How to partition or transform objects for caching
  in the proxy servers?
• Partitioning The leader blocks to reduce
  response time
• Hotspot The selected hotspot blocks to provide
  object preview
• Staging The blocks at peak data rates to reduce
  maximum WAN bandwidth
• Segment Partial object according to popularity
• Transcoding transform to low resolution object
  to reduce repeated processing time

6
Web Caching Cooperation

• Which proxy has the required object part?
• Hierarchical tree
• Parent proxy cache all objects in its children
• Large parent cache
• Directory based
• Each proxy keeps directory of cache contents in
  other cooperative proxy
• Update overheads
• Hash based
• Proxy no. Hash(object)
• Reorganize after proxy leaves

P
C
C
C
7
Cache Admission/Replacement Policies

• Which object parts should be kept in the caches?
• LRU removes the least recently accessed objects
  first.
• LFU removes the cold objects first.
• LRU-min, GD-size and LUV remove the large and
  least recently accessed objects first.
• Interval caching caches the short intervals in
  large video to guarantee continuous delivery for
  multimedia streams

8
Our new design

• Object partitioning
• Multiple different hotspots created at server
• different blocks from different cooperative proxy
• Full coverage of all blocks from cooperative
  proxy
• An object preview from each hotspot
• Cache Admission
• To maintain all blocks accessible from
  cooperative proxy servers
• Cooperation mechanism
• Message based
• Server control to maintain data security
• Proxy keeps full autonomy on its own cache
  content
• Limited number of cooperative proxy to maintain
  scalability

9
Multiple Hotspots

• E.g. Divide into low temporal resolution segments
• Group the blocks together to form multiple
  hotspots
• Every block belongs to at least one hotspot

1
2
3
4
5
6
7
8
9
10
11
12
Leader blocks
1
2
3
4
5
6
7
8
9
10
11
12
multiple hotspots
1
2
3
7
1
2
4
8
12
11
1
2
5
9
1
2
6
10
10
Cache Admission

• The hotspots are admitted to cache in priority
• uncached hotspots
• hotspots not cached in the regional network

A
S
C
B
D
11
Message Based Cooperation Algorithm

• Client sends request via local proxy to server S.
• Local proxy (LP) sends object request to server.
• Server S returns the decryption key and the list
  of coop proxy (CP) caching hotspots recently.
• LP sends request for a block to the nearest CP.
• CP returns a block or reject message to LP. If
  rejected, LP sends missing block request to
  server and server returns block to LP.
• LP returns the block to client and selects one
  hotspot to cache
• LP informs the server about which hotspot is
  cached.
• Server updates the list of coop proxy with
  hotspots.

12
Simulation Parameters

• 5 servers, 1000 video titles per server
• Video length uniformly distributed with mean 125
  blocks
• 10 hotspots per video
• Zipf-like popularity c/i(1-?), ?0.271
• 50 proxy servers
• proxy cache size 25,000 blocks
• Cache replacement function 1/((T-T)log(blockno
  16))
• Server returns list of 4 coop proxy per hotspots
• Network latency exponentially distributed with
  different mean value local, regional remote

13
Simulation

• Comparing methods
• cooperative hotspot,
• non-cooperative hotspot and
• variable length segment based
• Compare
• Byte hit ratio
• Stream service time
• Number of server streams
• Number of proxy streams
• Stream response time

14
Increase the byte hit ratio by 70 using coop
proxy cache Best when all hotspots can be found
in regional coop proxy
15
Reduce service time with more cooperative proxy
returned per hotspot
16
The lowest server load. Number of server streams
increases with number of requesting proxy servers
at the lowest rate
17
Limited increase in proxy load for cooperation.
Smallest increase when the no. of proxy in region
no. of hotspots
18
Reduce service time, more efficient under light
load condition Only lightly loaded proxy should
participate.
19
Low response time similar to non-cooperative
hotspot under light load. Reject cooperative
requests under heavy load.
20
Summary

• Cooperative proxy with multiple hotspots
• Provide byte hits for 70 of data from previous
  compulsory cache misses
• Reduce service time of streams
• Dynamically balance loads from hot servers to
  lightly loaded cooperative proxy servers
• The method is best when
• All hotspots are cached in regional coop proxy
• Number of hotspots number of proxy in the
  region
• The server returns more cooperative proxy

21
Work to continue

• Scalability issue increase in both number of
  servers and proxy servers
• Region based cooperation
• Detailed performance model
• Enhancements
• Variable length hotspots
• Content discovery using hints
• Guarantee server streams without reservation