An End-to-end Architecture for Quality-Adaptive Streaming Applications in Best-effort Networks

1
An End-to-end Architecture for Quality-Adaptive
Streaming Applications in Best-effort Networks

• Reza Rejaie
• reza_at_isi.edu
• USC/ISI
• http//netweb.usc.edu/reza
• April 13, 1999

2
Motivation

• Rapid growth in deployment of realtime
  streams(audio/video) over the Internet
• TCP is inappropriate for realtime streams
• The Internet requires end-system to react to
  congestion properly and promptly
• Streaming applications require sustained
  consumption rate to deliver acceptable and stable
  quality

3
Best-effort Networks (The Internet)

• Shared environment
• Bandwidth is not known a prior
• Bandwidth changes during a session
• Seemingly-random losses
• TCP-based traffic dominates
• End-to-end congestion control is crucial for
  stability, fairness high utilization
• End-to-end congestion control in a TCP-friendly
  fashion is the main requirement in the Internet

4
Streaming Applications

• Delay-sensitive
• Semi-reliable
• Rate-based
• Require QoS from the end-to-end point of view

Encoder
Adaptation
Source
Server
TCP
TCP
Internet
Buffer
Decoder
Display
5
The Problem

• Designing an end-to-end congestion control
  mechanism
• Delivering acceptable and stable quality while
  performing congestion control

6
Outline

• The End-to-end Architecture
• Congestion Control (The RAP protocol)
• Quality Adaptation
• Extending the Architecture
• Multimedia Proxy Caching
• Contributions
• Future Directions

7
The End-to-end Architecture
Error Control
Quality Adaptation
Cong. Control
Acker
Playback Buffer
Internet
Buffer Manager
Buffer Manager
Transmission Buffer
Decoder
Archive
Adaptation Buffer
Server
Client
Data path
Control path
8
Outline

• The End-to-end Architecture
• Congestion Control (The RAP Protocol)
• Quality Adaptation
• Extending the Architecture
• Multimedia Proxy Caching
• Contributions
• Future Directions

9
Previous works on Congestion Ctrl.

• Modified TCP
• Jacob et al. 97, SCPCen et al. 98
• TCP equation
• Mathis et al. 97, Padhye et al. 98
• Additive Inc., Multiplicative Dec.
• LDASisalem et al. 98
• NETBLTLixia!
• Challenge TCP is a moving target

10
Overview of RAP

• Decision Function
• Increase/Decrease Algorithm
• Decision Frequency
• Goal to be TCP-friendly

Decision Function
Rate
Increase/Decrease Algorithm
--

Time
Decision Frequency
11
Congestion Control Mechanism

• Adjust the rate once per round-trip-time (RTT)
• Increase the rate periodically if no congestion
• Decrease the rate when congestion occurs
• Packet loss signals congestion
• Cluster Loss
• Grouping losses per congestion event

12
Rate Adaptation Algorithm

• Coarse-grain rate adaptation
• Additive Increase, Multiplicative Decrease (AIMD)
• Extensive simulations revealed
• TCPs behavior substantially varies with network
  conditions, e.g. retransmission timeout, bursty
• TCP is responsive to a transient congestion
• AIMD only emulates window adjustment in TCP

13
Rate Adaptation Algorithm(contd)

• Fine-grain rate adaptation
• The ratio of short-term to long-term average RTT
• Emulates ACK-clocking in TCP
• Increase responsiveness to transient congestion

14
Coarse vs fine grain RAP fig
Impact of fine-grain rate adaptation
15
RAP Simulation
TCP Traffic

• RAP against Tahoe, Reno, NewReno SACK
• Inter-dependency among parameters
• Config. parameters
• Bandwidth per flow
• RTT
• Number of flows

TCP Sinks
TCP Sources
SW
SW
RAP Sinks
RAP Sources
Avg. RAP BW
Fairness Ratio
Avg. TCP BW
RAP Traffic
16
Fairness ratio across the parameter space without
F.G. adaptation
17
Fairness ratio across the parameter space with
F.G. adaptation
18
Impact of RED switches on Fairness ratio
19
Summary of RAP Simulations

• RAP achieves TCP-friendliness over a wide range
• Fine grain rate adaptation extends inter-protocol
  fairness to a wider range
• Occasional unfairness against TCP traffic is
  mainly due to divergence of TCP congestion
  control from AIMD
• Pronounced more clearly for Reno and Tahoe
• The bigger TCPs congestion window, the closer
  its behavior to AIMD
• RED gateways can improve inter-protocol sharing
• Depending on how well RED is configured
• RAP is a TCP-friendly congestion controlled UDP

20
Outline

• The End-to-end Architecture
• Congestion Control (The RAP protocol)
• Quality Adaptation
• Extending the Architecture
• Multimedia Proxy Caching
• Contributions
• Future Directions

21
Quality Adaptation
Error Control
Quality Adaptation
Cong. Control
Acker
Playback Buffer
Internet
Buffer Manager
Buffer Manager
Transmission Buffer
Decoder
Archive
Adaptation Buffer
Server
Client
Data path
Control path
22
The Problem

• Delivering acceptable and stable quality while
  performing congestion control
• Seemingly random losses result in random
  potentially wide variations in bandwidth
• Streaming applications are rate-based

23
Role of Quality Adaptation

• Buffering only absorb short-term variations
• Long-lived session could result in buffer
  overflow or underflow
• Quality Adaptation is complementary for buffering
• Adjust the quality with long-term variations in
  bandwidth

BW(t)
Time
24
Mechanisms to Adjust Quality

• Adaptive encoding Ortega 95, Tan 98
• CPU-intensive
• Switching between multiple encoding
• High storage requirement
• Layered encodingMcCanne 96, Lee 98
• Inter-layer decoding dependency
• When/How much to adjust the quality?

25
Assumptions Goals

• Assumptions
• AIMD variations in bandwidth(rate)
• Linear layered encoding
• Constraint
• Obeying congestion controlled rate limit
• Goal
• To control the level of smoothing

26
Layered Quality Adaptation
bw (t)
2
C
buf
2
bw (t)
bw (t)
2
Layer 2
1
BW(t)
BW(t)
C

buf
Internet
1
bw (t)
bw (t)
1
Layer 1
0
C
Display
buf
0
bw (t)
0
Layer 0
Decoder
Filling Phase
Quality Adaptation
Draining Phase
BW(t)
Linear layered stream
a
c
C
BW(t)
Consumption rate
C
b
C
Time(msec)
Time(sec)
27
Buffering Tradeoff
bw (t)
2
C
buf
2

• Each buffering layer can only contribute at most
  C(bps)
• Buffering for more layers provides higher
  stability

bw (t)
1
C
BW(t)
buf
1
bw (t)
0
C
buf
0

• Buffered data for a dropped layer is useless for
  recovery
• Buffering for lower layers is more efficient

BW(t)

• What is the optimal buffer distribution for a
  single back-off scenario?

nC
Time
28
Optimal Inter-layer Buffer Allocation
Draining Phase
Filling Phase
BW(t)

• Optimal buffer state depends on time of the
  back-off
• Draining pattern depends on the buffer state
• Back-off occurs randomly
• Keep the buffer state as close to the optimal as
  possible during the filling phase

C
C
4C
Time
Buf. data
BW share of L0
Buf. data
BW share of L1
BW share of L2
Buf. data
29
Adding Dropping
BW(t)

• Add a layer when buffering is sufficient for a
  single back-off
• Drop a layer when buffering is insufficient for
  recovery
• Random losses could result in frequent add and
  drop
• unstable quality
• Conservative adding results in smooth changes in
  quality

Time
Buf. data for L0
Buf. data for L1
Buf. data for L2
30
Smoothing

• Conservative adding
• When average bandwidth is sufficient
• When sufficient buffering for K back-offs
• Buffer constraint is preferred and sufficient
• Directly relate time of adding to the buffer
  state
• Effectively utilizes the available bandwidth
• K is a smoothing factor
• Short-term quality vs long-term smoothing

31
Smooth Filling Draining
Proper Buf. State recovery from 1 backoff
Proper Buf. State recovery from 2 backoffs
Proper Buf. State recovery from K backoffs
Add a Layer
Drop a Layer
Filling
Draining
32
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)
33
Adapting to network load
(K 4)
KB/s
TX rate Quality
C 10
90 Time(sec)
30
60
KB
17.5
Buf. L3(KB)
17.5
Buf. L3(KB)
17.5
Buf. L3(KB)
17.5
Buf. L3(KB)
30
60
90 Time(sec)
34
No of Dropped Layers
35
Summary of the QA results

• Quality adaptation mechanism can efficiently
  control the quality
• Smoothing factor allows the server to trade
  short-term improvement with long-term smoothing
• Buffer requirement is low
• Deploying for live but non-interactive sessions!

36
Limitation of the E2E Approach

• Delivered quality is limited to the average
  bandwidth between the server and client
• Solutions
• Mirror servers
• Multimedia proxy caching

Client
Client
Client
Internet
Server
Quality(layer)
Time
37
Outline

• The End-to-end Architecture
• Congestion Control (The RAP protocol)
• Quality Adaptation
• Extending the Architecture
• Multimedia Proxy Caching
• Contributions
• Future Directions

38
Multimedia Proxy Caching

• Assumptions
• Proxy can perform
• End-to-end congestion ctrl
• Quality Adaptation
• Goals
• Improve delivered quality
• Low-latency VCR-functions
• Natural benefits of caching

Client
Client
Client
Proxy
Internet
Server
39
Challenge

• Cached streams have variable quality
• Layered organization provides opportunity for
  adjusting the quality

L
4
Quality (layer)
L
3
L
2
L
1
L
0
Time
40
Issues

• Delivery procedure
• Relaying on a cache miss
• Pre-fetching on a cache hit
• Replacement algorithm
• Determining popularity
• Replacement pattern

41
Cache Miss Scenario
Client
Client
Client

• Stream is located at the original server
• Playback from the server through the proxy
• Proxy intercepts and caches the stream
• No benefit in a miss scenario

Proxy
Internet
Server
42
Cache Hit Scenario
Client
Client
Client

• Playback from the proxy cache
• Lower latency
• May have better quality!
• Available bandwidth allows
• Lower quality playback
• Higher quality playback

Proxy
Internet
Server
43
Lower quality playback

• Missing pieces of the active layers are
  pre-fetched on-demand
• Required pieces are identified by QA
• Results in smoothing

L
4
L
Quality (no. active layers)
3
L
2
L
1
L
0
Time
44
Higher quality playback

• Pre-fetch higher layers on-demand
• Pre-fetched data is always cached
• Must pre-fetch a missing piece before its
  playback time
• Tradeoff

L
4
L
Quality (no. active layers)
3
L
2
L
1
L
0
Time
45
Replacement Algorithm

• Goal converge the cache state to optimal
• Average quality of a cached stream depends on
• popularity
• average bandwidth between proxy and recent
  interested clients
• Variation in quality inversely depends on
• popularity

Client
Client
Client
Proxy
Internet
Server
46
Popularity

• Number of hits during an interval
• Users level of interest (including
  VCR-functions)
• Potential value of a layer for quality adaptation
• Calculate whit on a per-layer basis
• Layered encoding guarantees monotonically
  decrease in popularity of layers

whit PlaybackTime(sec)/StreamLength(sec)
47
Replacement Pattern

• Multi-valued replacement decision for multimedia
  object
• Coarse-grain flushing
• on a per-layer basis
• Fine-grain flushing
• on a per-segment basis

Cached segment
Fine-grain
Quality(Layer)
Coarse-grain
Time
48
Summary of Multimedia Caching

• Exploited characteristics of multimedia objs
• Proxy caching mechanism for multimedia streams
• Pre-fetching
• Replacement algorithm
• Adaptively converges state of the cache to the
  optimal

49
Contributions

• End-to-end architecture for delivery of
  quality-adaptive multimedia streams
• RAP, a TCP-friendly cong. ctrl mechanism over a
  wide range of network conditions
• Quality adaptation mechanism that adjusts the
  delivered quality with a desired degree of
  smoothing
• Proxy caching mechanism for multimedia streams to
  effectively improve the delivered quality of
  popular streams

50
Future Directions

• End-to-end Congestion Control
• RAPs behavior in the presence web-like traffic
• Emulating timer-driven regime TCP
• Bi-directional RAP connections, Reverse ns
  forward path congestion control
• Experiments over CAIRN the Internet
• Integration of RAP and congestion manager
• Adopting RAP into class-based QoS
• Using RAP for multicast congestion control
• Congestion control over wireless networks

51
Future Directions(contd)

• Quality Adaptation
• Extending to other rate adaptation mechanisms
• Multimedia Proxy Caching
• Other replacement patterns popularity
  functions(e.g. chunk-based)
• Traffic Measurement and Characterization
• Imiprical evaluation of streaming applications

52
An End-to-end Architecture for Quality-Adaptive
Streaming Applications in Best-effort Networks

• Reza Rejaie
• reza_at_isi.edu
• USC/ISI
• http//netweb.usc.edu/reza
• April 7, 1999

53
Thank you
Reza Rejaie
reza_at_isi.edu
http//netweb.usc.edu/reza
54
Target Environment
TCP Traffic
TCP Traffic
55
Optimal Buffer Allocation
Optimal buffer state is not unique S1 and S2 are
extreme cases S1 requires more buffering
layers S2 requires more buffer share per
layer Buffer allocation for S1 can recover from
S2 but not vice versa
Backoff 2
Scenario 1
Scenario 2
Scenario 3