Embed Size (px)
Citation preview
1
NUS.SOC.CS5248Ooi Wei Tsang
Proxy Caching for Streaming Media
NUS.SOC.CS5248Ooi Wei Tsang
2
You Are Here
Network
Encoder
Sender
Middlebox
Receiver
Decoder
NUS.SOC.CS5248Ooi Wei Tsang
3
Cache Proxies for Web
A
NUS.SOC.CS5248Ooi Wei Tsang
4
Hierarchical Caching
A
B
NUS.SOC.CS5248Ooi Wei Tsang
5
Cooperative Caching
A B
NUS.SOC.CS5248Ooi Wei Tsang
6
Distributed Caching
A B
NUS.SOC.CS5248Ooi Wei Tsang
7
Streaming Media vs. Webpage
NUS.SOC.CS5248Ooi Wei Tsang
8
Video Access Pattern
by S. Acharya and B. Smith in 1999
Study at Lulea University, Sweden 55% complete, 45% stop very early High temporal locality
NUS.SOC.CS5248Ooi Wei Tsang
9
Prefix Access Distribution
10 30 50 70 90
% Played
NUS.SOC.CS5248Ooi Wei Tsang
10
Video Popularity
Assume Zipf “Law”Probability of access to i-th most popular video is
11
NUS.SOC.CS5248Ooi Wei Tsang
Benefits of Caching
NUS.SOC.CS5248Ooi Wei Tsang
12
Reduce Access Latency
:) :(
NUS.SOC.CS5248Ooi Wei Tsang
13
Reduce Server Load
NUS.SOC.CS5248Ooi Wei Tsang
14
Reduce Start-up Latency
NUS.SOC.CS5248Ooi Wei Tsang
15
Hide Network Congestion
NUS.SOC.CS5248Ooi Wei Tsang
16
Other Issues
What to cache?Who to fetch from?When cache is full, who to kick out?
How to measure popularity?Can cache adapt to popularity?
17
NUS.SOC.CS5248Ooi Wei Tsang
What to Cache?
NUS.SOC.CS5248Ooi Wei Tsang
18
Segmentation
Cache “all or none” is badDivide media file into segments and consider each segment individually
NUS.SOC.CS5248Ooi Wei Tsang
19
Effects of Size S
Large S : Low utilizationSmall S : Lots of gaps
NUS.SOC.CS5248Ooi Wei Tsang
20
Prefix Caching Policy
1 Chunk = k segments
NUS.SOC.CS5248Ooi Wei Tsang
21
Caching Policy
Basic unit of caching: segmentCache prefix in chunkReplace suffix in chunkNever replace segments in curr chunk
22
NUS.SOC.CS5248Ooi Wei Tsang
Where To Fetch From?
NUS.SOC.CS5248Ooi Wei Tsang
23
Cooperative Caching
A B
NUS.SOC.CS5248Ooi Wei Tsang
24
Fetch from server
B
A
Server
Client 2Client 1
NUS.SOC.CS5248Ooi Wei Tsang
25
Fetch from fellow proxy
B
A
Server
Client 2Client 1
NUS.SOC.CS5248Ooi Wei Tsang
26
Issues
How to advertise?How to choose “helper”?
NUS.SOC.CS5248Ooi Wei Tsang
27
How to Advertise?
Balance betweennetwork loadfreshness of information
NUS.SOC.CS5248Ooi Wei Tsang
28
Scalable Advertisement
Expanding Ring Advertisement
16 132 264 4128 8
TTL PERIOD
NUS.SOC.CS5248Ooi Wei Tsang
29
How to Choose Helper?
Consideration for Static Cache network distance (1,2,3,4) number of streams being served avoid frequent switches
Build a cost function, integrating the metrics
NUS.SOC.CS5248Ooi Wei Tsang
30
Cost Function
Cost for retrieving a segment from node X to node Y=
NUS.SOC.CS5248Ooi Wei Tsang
31
Algorithm
Consider the next gap in local caches
Find the next helper with minimum cost, and can fill in at least k segments in gaps
32
NUS.SOC.CS5248Ooi Wei Tsang
Distributed Caching
Y. Chae et. al.JSAC 2000
NUS.SOC.CS5248Ooi Wei Tsang
33
Cooperative vs DistributedCooperative caching caches independently
Distributed caching caches as a team
NUS.SOC.CS5248Ooi Wei Tsang
34
Cold Start
B
A
Server
new clip!
NUS.SOC.CS5248Ooi Wei Tsang
35
Segment Map
Local segment mapWhich segment I should cache?
Global segment mapWho is suppose to cache what?
NUS.SOC.CS5248Ooi Wei Tsang
36
Cache Hit
B
A
Server
NUS.SOC.CS5248Ooi Wei Tsang
37
Cache Miss
B
A
Server
NUS.SOC.CS5248Ooi Wei Tsang
38
Distributed Caching
B
A
Server
NUS.SOC.CS5248Ooi Wei Tsang
39
Problems
Who should cache what?
Which segment to kick out?
How to adapt segment distribution?
NUS.SOC.CS5248Ooi Wei Tsang
40
Who should cache what?
RCache schemeSegment video into equal size segments
A proxy will cache each segment with some probability
NUS.SOC.CS5248Ooi Wei Tsang
41
RCache
Np proxies
video of length Lv
divide into Ns equal segments
Each proxy caches each segment with a/Np probability
NUS.SOC.CS5248Ooi Wei Tsang
42
Analysis
Probability that whole video is cached is
Ns:: num of segmentsNp: num of proxiesa/Np: prob of caching 1 segment
NUS.SOC.CS5248Ooi Wei Tsang
43
RCache’s Segmentation
Video is divided into segments of equal length
Can we do better?
NUS.SOC.CS5248Ooi Wei Tsang
44
Bimodal Distribution
10 30 50 70 90
% Played
NUS.SOC.CS5248Ooi Wei Tsang
45
Silo
segment size probabilityof storage
NUS.SOC.CS5248Ooi Wei Tsang
46
Further Improvement
segment size probabilityof storage
NUS.SOC.CS5248Ooi Wei Tsang
47
Problems
Who should cache what?
Which segment to kick out?
How to redistribute data?
NUS.SOC.CS5248Ooi Wei Tsang
48
Segment “Popularity”
For each video iFor each segment j
F(i,j) = Prob(i is accessed)*Prob(j is accessed)
NUS.SOC.CS5248Ooi Wei Tsang
49
Rainbow Algorithm
less popular
NUS.SOC.CS5248Ooi Wei Tsang
50
Rainbow Algorithm
Problem: Many large video – too many segments
computationally expensive to sort
Solution: Just approximate by quantizing popularity
NUS.SOC.CS5248Ooi Wei Tsang
51
Rainbow Algorithm
less popular
NUS.SOC.CS5248Ooi Wei Tsang
52
Problems
Who should cache what?
Which segment to kick out?
How to redistribute data?
NUS.SOC.CS5248Ooi Wei Tsang
53
Data Redistribution
When popularity changes, need to redistribute.
Redistribute “on-demand” (lazy)
NUS.SOC.CS5248Ooi Wei Tsang
54
Cache Token
Each segment have two bits(T,C)
T: I am suppose to have the segment
C: I have the segment
NUS.SOC.CS5248Ooi Wei Tsang
55
Data Redistribution
(T=1,C=1)(T=0,C=0)(T=1,C=0) (T=0,C=1)
NUS.SOC.CS5248Ooi Wei Tsang
56
Example
I J
A CB D
old
new
T=0,C=1 T=0,C=1
T=1,C=0 T=1,C=0 T=1,C=0 T=1,C=0
NUS.SOC.CS5248Ooi Wei Tsang
57
Example
I J
A CB D
old
new
T=1,C=1 T=0,C=1
T=0,C=0 T=1,C=0 T=1,C=0 T=1,C=0
NUS.SOC.CS5248Ooi Wei Tsang
58
Example
I J
A CB D
old
new
T=1,C=1 T=1,C=1
T=0,C=0 T=1,C=0 T=1,C=0 T=0,C=0
NUS.SOC.CS5248Ooi Wei Tsang
59
Example
I J
A CB D
old
new
T=1,C=1 T=1,C=1
T=0,C=0 T=1,C=1 T=1,C=0 T=0,C=0
60
NUS.SOC.CS5248Ooi Wei Tsang
Quality Adaptive Caching
Reza Rajaie et. al.INFOCOM 2000
NUS.SOC.CS5248Ooi Wei Tsang
61
Scenario (10am)
A
Server
Client 2
NUS.SOC.CS5248Ooi Wei Tsang
62
Scenario (1am)
A
Server
Client 2
NUS.SOC.CS5248Ooi Wei Tsang
63
Problem:Cache interfere with congestion control algorithm
Solution:Make cache aware of quality adaptation
NUS.SOC.CS5248Ooi Wei Tsang
64
Solutions
Making cache “quality-aware”PrefetchReplacement Algorithm
NUS.SOC.CS5248Ooi Wei Tsang
65
Cache Miss
A
Server
Client 2
NUS.SOC.CS5248Ooi Wei Tsang
66
Cache Hit
A
Server
Client 2
repair +prefetch
NUS.SOC.CS5248Ooi Wei Tsang
67
Peeking Inside the Cache
NUS.SOC.CS5248Ooi Wei Tsang
68
Cache Hit: Repair
NUS.SOC.CS5248Ooi Wei Tsang
69
Cache Hit: Prefetch
NUS.SOC.CS5248Ooi Wei Tsang
70
Prefetch Algorithm
prefetchwindow
playbackpoint
NUS.SOC.CS5248Ooi Wei Tsang
71
Proxy Request to Server
Multiple requests (for different clients) are batched.
NUS.SOC.CS5248Ooi Wei Tsang
72
Server Response
Missing segments are sent in decreasing priority
1
2
3 4
NUS.SOC.CS5248Ooi Wei Tsang
73
Server Response
Send as many segments as possible until next prefetch request
NUS.SOC.CS5248Ooi Wei Tsang
74
Trade-offs
How far in the future should we prefetch?
NUS.SOC.CS5248Ooi Wei Tsang
75
Solutions
Making cache “quality-aware”PrefetchReplacement Algorithm
NUS.SOC.CS5248Ooi Wei Tsang
76
Goal: converge to efficient state
if a stream is popularaverage quality is highvariation in quality is low
Goal of Replacement
NUS.SOC.CS5248Ooi Wei Tsang
77
The Algorithm
NUS.SOC.CS5248Ooi Wei Tsang
78
Trashing and Locking
NUS.SOC.CS5248Ooi Wei Tsang
79
Choosing Victim
whit (weighted hit) =Tplay/Ttotal
Calculate whit for each layer in a stream over a popularity window
NUS.SOC.CS5248Ooi Wei Tsang
80
Example
WHIT Lock? Name Layer
5.9 1 Nemo 0
4.3 1 Nemo 1
4.0 0 Matrix 2 0
3.9 0 Matrix 2 1
1.1 0 Gigli 0
NUS.SOC.CS5248Ooi Wei Tsang
81
Static Cache
Cache segments in proxy do not changed over time
Can we exploit further properties of streaming media?
82
NUS.SOC.CS5248Ooi Wei Tsang
Dynamic Caching
NUS.SOC.CS5248Ooi Wei Tsang
83
Motivating Scenario
t = 0, R1 requests for stream M
t = , R2 requests for stream M
Ideally, R1 and R2 should share a multicast of M
NUS.SOC.CS5248Ooi Wei Tsang
84
needs to be“patched”
share withR1 from cache
NUS.SOC.CS5248Ooi Wei Tsang
85
Using Dynamic Cache
R2 request stream MProxy allocate a ring bufferCache the most recent -seconds of M sent to R1
R2 get prefix of M from other places, and rest from proxy
NUS.SOC.CS5248Ooi Wei Tsang
86
Extending to N Receivers
