Reza Rejaie

1
Design Evaluations of Multimedia Proxy Caching
Mechanisms for the Internet

• Reza Rejaie
• ATT Labs Research
• Menlo Park
• http//www.research.att.com/reza
• In collaboration with
• Mark Handley(ACIRI at ICSI)
• Deborah Estrin and Haobo Yu (ISI/USC)

2
Motivation

• Rapid growth in client-server media (audio/video)
  streaming over the Internet
• Delivered quality in client-server streaming
  could be low and unstable
• Deployment of client-server streaming in a large
  scale imposes a heavy load on the network

3
Goal

• Supporting unicast multimedia playback over the
  Internet
• High quality
• Large scale availability
• Low-latency VCR-functions
• Efficient async. access

4
The Internet

• Todays Internet
• Best-effort service
• Resources are shared with TCP-traffic
• Unpredictable changes in available bandwidth
• Tomorrows Internet
• Different classes of service
• Resources are shared within a class
• Internet applications should be rate-adaptive
• Fairness Stability

5
Client-server Internet Streaming

• Internet streaming applications should be quality
  adaptive(QA)
• QA is often needed in a shared environment, e.g.
• Diff-serve without per-flow admission control
• Shared reservation
• Streaming over best-effort class
• Quality adaptation
• Adjust the quality with long-term changes in BW
• Limited quality Limited scalability

6
Client-Server Streaming An Example
Source
Encoder
Adaptation
Server
TCP
TCP
Internet
Buffer
Decoder
Display
7
The Problem

• Designing an efficient effective QA mechanism
  for client-server streaming?
• Rapid changes in quality are disturbing
• Accommodating large-scale access to multimedia
  content efficiently?
• Focus on unicast playback

8
Overview

• A client-server architecture
• Layered quality adaptation
• Extending the architecture
• Replication or Proxy Caching?
• Multimedia proxy caching (MCaching)
• Evaluation space
• MCaching design issues
• Pre-fetching
• Replacement
• Simulation results

9
Mechanisms to Adjust Quality

• Adaptive encodingOrtega95, Tan98
• CPU-intensive
• Switching between multiple encoding
• High storage requirement
• Layered encodingMcCanne96, Lee98
• Inter-layer decoding dependency

10
Layered QA The Idea

• Fine-grained adaptation
• short-term variations in bandwidth are absorbed
  by receiver buffering
• Coarse-grained adaptation
• Long-term variations in bandwidth result in layer
  add/drop(i.e. quality adjustment)
• Inter-layer buffer distribution is adjusted by
  inter-layer bw distribution

11
Client-Server Architecture
Decoder
bw (t)
Quality Adaptation
2
C
buf
2
bw (t)
ACK
bw (t)
2
BW(t)
Layer 2
1
BW(t)
C
Internet

buf
1
Acker
Rate Adaptation
bw (t)
bw (t)
1
Layer 1
0
C
buf
Data
0
bw (t)
0
Layer 0
BW(t)
Linear layered stream
Error Control
Display
12
Effect of smoothing factor
(K 2)
KB/s
TX rate Quality
C 10
40 Time(sec)
Buf. L3(KB)
9.5
9.5
Buf. L2(KB)
9.5
Buf. L1(KB)
9.5
Buf. L0(KB)
40 Time(sec)
(K 4)
KB/s
TX rate Quality
C 10
40 Time(sec)
17.5
Buf. L3(KB)
Buf. L2(KB)
17.5
17.5
Buf. L1(KB)
17.5
Buf. L0(KB)
40 Time(sec)
13
Issues with E2E Streaming

• Limited quality
• Limited scalability
• Load on the server/network
• Asynchronous access could be inefficient
• RTT could be high
• High delay VCR-functions
• Large startup delay
• Content should be close to interested clients

14
Extending the Architecture

• Replications
• Proxy Caching
• Similar in nature
• Only applicable to playback streaming
• These mechanisms are complementary

15
Replication
Campus
ISP
Internet
16
Proxy Caching
Campus
ISP
Internet
17
Replication vs Proxy Caching

• Location Proxy is often closer to clients
• No bottleneck
• Min RTT
• Higher locality of reference!
• Content management
• A proxy adaptively caches popular streams from
  different servers based on clients interest
• A mirror server statically replicates content of
  a single server
• Administration
• Replication Proxy Caching are complementary

18
Multimedia Proxy Caching (MCaching)

• Assuming locality of reference exists
• Goal Cache popular streams with appropriate
  quality
• Load on the network/server could be minimized
• Delivered quality could be maximized
• Async. access can be supported efficiently
• Low-latency VCR-functions
• Appropriate quality is determined by
• Client bandwidth
• Popularity of streams
• Max. deliverable quality is not guaranteed
• Caching appropriate quality Higher performance

19
Extending Existing Web Caches?

• Notion of quality for cached objects does not
  exist in Web caching
• Atomic delivery
• Atomic replacement
• Existing multimedia caches
• A web cache a media player
• Two possible extensions
• LQC Cache and replay low-quality streams
• HQC Pre-fetch max. quality of poplar streams
• Existing Web caches can not support Mcaching
  efficiently

20
Performance Evaluation

• Goal of Web caching
• to maximize Byte-Hit-Ratio
• Goal of MCaching
• to maximize Byte-Hit-Ratio, and
• to maximize delivered quality
• MCaching performance should be evaluated across
  quality-BHR plan

21
MCaching Evaluation Space
Maximum Quality
I
HQC
Quality
I
Max-deliverable Quality
O
O
MC
I
O
Bottleneck Quality
X
X
X
LQC
BHR
22
MCaching The idea

• Exploit layered organization
• Relay on a cache miss
• Pre-fetch on a cache hit
• If higher quality is required
• Server should support fine-grained access
• Proxy is able to properly glue intranet
  Internet transport

Client
Client
Client
Proxy
Internet
Server
23
MCaching Design Issues

• Fine-grained pre-fetching
• Online vs offline
• Timing efficiency
• Sharing pre-fetching BW among concurrent sessions
• Fine-grained replacement
• What granularity is appropriate?
• Fine-grained popularity, e.g. per-segment
• Interactions between these mechanisms should
  properly adjust quality of cached st.

24
Pre-fetching An Example

• Missing pieces of the active layers are
  pre-fetched on-demand
• Required pieces are identified by QA
• Pre-fetching results in improvement of quality
• Pre-fetched data is always cached

L
4
L
Quality (no. active layers)
3
L
2
L
1
L
0
Time
25
Fine-grained Online Pre-fetching
Pre-fetch
Playback

• Pre-fetching stream is congestion controlled
• Gradual improvement
• Pre-fetching can repair losses
• Pre-fetching playback should remain loosely
  sync.
• Trade-off (T)
• Sliding-window
• Batch of missing segments
• Prioritized delivery
• Should also incorporate utility of a segment

Client
Proxy
Server
Pre-fetching Window
L
4
Quality (layers)
L
3
L
2
L
1
L
0
Time
T
W
t
p
A Segment
26
Fine-grained Replacement

• Allows fine-grained quality adjustment
• Avoids fragmentation of cache space
• Requires fine-grained popularity

27
Popularity

• Clients level of interest (e.g. VCR-functions)
• Number of hits during a recent window
• Pipelining - Hit could be any value within
  0..1
• Clients BW Potential usage of a layer for
  quality adaptation during future sessions
• Calculate whit on a per-layer basis
• Layered encoding guarantees monotonically
  decrease in popularity of layers

whit weighted_hit PlaybackTime(sec)/StreamLeng
th(sec)
28
Replacement Pattern

• Per-layer popularity
• Victim layer
• Granularity of replacement
• On a per-segment basis
• Demand-driven
• Victim segment?
• Utility of a segment?
• Locking is needed

Cached segment
Fine-grained
Coarse-grained
Quality(Layer)
Time
29
Simulation Setup
HTTP-like Traffic

• No metric for quality ?
• Sequential delivery
• 10 streams with uniform random dist. 1..10min
• Cache size x of data set
• Evaluation space
• Effect of client bandwidth
• Distribution of requests between two clients
• Effect of popularity
• Stream popularity conforms to the Zipfs law
• Per-flow vs Aggregate

bw
1
Client 1
Proxy
Server
bw
bw
sp
Client 2
2
bw_sp bw2 bw_sp 56 Kbps bw1
1.5 Mbps bw2 56 Kbps
30
Quality of Cached Streams
Per-layer completeness(most popular stream)
Per-layer completeness ()
Time(sec)
31
Quality vs BHR
Average Cached Quality(layer)
Byte-Hit-Ratio
32
Quality vs Network Load
Average Cached Quality(layer)
Network Load
33
Status

• Prototyping a modular client-server architecture
  for Internet streaming
• Cong. Ctrl Quality Adaptation Error Ctrl
• Layered encoding
• Prototyping a MCache on top of Squid

34
Future Directions

• Given utility function for a (encoding, content),
  can we optimize
• Quality adaptation
• Quality adaptation Error control
• Fine-grained replacement pre-fetching
• Multimedia traffic measurement and
  characterization