Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
A Case for a
Coordinated Video
Control Plane
Xi Liu, Florin Dobrian, Henry Milner,
Junchen Jiang, Vyas Sekar, Ion Stoica,
Hui Zhang
(Conviva, CMU, Intel, and UC Berkeley)
Video Is Dominating the Internet Traffic
Netflix traffic alone exceeds 20% of US traffic1
2011’s Cisco Visual Networking Index2
2011: video represents 51% of the Internet traffic
2016: all types of video will represent 86% of the Internet traffic
The Internet is becoming a Video Network
2http://web.cs.wpi.edu/~claypool/mmsys-2011/Keynote02.pdf
1http://blogs.cisco.com/sp/comments/cisco\_visual\_networking\_index\_forecast\_annual\_update
Video Ecosystem: Data-Plane
Video Source
Encoders & Video Servers
CMS and Hosting
Content Delivery Networks (CDNs)
ISP & Home Net
Screen
Video Player
Video Quality Matters [Sigcomm’11]
Quality has substantial impact on viewer engagement
Need to ensure uninterrupted streaming at high bitrates
Buffering ratio is most critical across video traffic types
Highest impact for live: 1% of buffering reduced play time by 3min
1% increase in buffering can lead to more than 60% loss in
audience over one month
Our Argument
CDN performance varies widely in time, geography, and
ISPs
Opportunity for significantly improving video Quality by
selecting best CDN (and bitrate) for each viewer
Hence, we argue for a logically centralized control plane to
dynamically select CDN and bitrate
Assumptions:• Content is encoded at multiple bitrates• Content is delivered by multiple CDNs
How do We Collect Data?
Streaming Module
UI Controller
Content Manager
Messaging & Serialization
Tobackend
HTTPS
Automatic Monitoring
Player Insight
Pla
yer
Ap
plic
atio
n
Automatic and continuous monitoring of video player
Flash: NetStream, VideoElement
Silverlight: MediaElement, SmoothStreamMediaElement
iOS: MPMoviePlayerElement
What Traffic do We See?
Close to two billions streams per month
Mostly premium content providers (e.g., HBO, ESPN, Disney) but also
User Generated Video sites (e.g., Ustream)
Live events (e.g., NCAA March Madness, FIFA World Cup, MLB), short
VoDs (e.g., MSNBC), and long VoDs (e.g., HBO, Hulu)
Various streaming protocols (e.g., Flash, SmoothStreaming, HLS), and
devices (e.g., PC, iOS devices, Roku, XBOX, …)
Traffic from all major CDNs, including ISP CDNs (e.g., Verizon, AT&T)
CDN Performance Varies Widely
CDNs Vary in Performance over Geographies and Time
There is no single best CDN across geographies,
network, and time
25%
50%
25%
CDN 1
CDN 2
CDN 3
• Metric: buffering ratio
• One month aggregated data-set– Multiple Flash (RTMP) customers
– Three major CDNs
• 31,744 DMA-ASN-hour with > 100 streams from each CDN– DMA: Designated Market Area
• Percentage of DMA-ASN-hour partitions a CDN has lowest buffering ratio
Washington DC (Hagerstown):
ASN-CXA-ALL
10% 20% 100%30% 40% 50% 60% 70% 80% 90%
Washington, DC viewer experience
differed greatly… Comcast viewers got the best
streams from CDN 1 51% of the time
and only 9% from CDN 2
Washington DC (Hagerstown):
VZGNI-TRANSIT (19262)
Verizon users got the best streams from CDN 1 only
17% of the time and 77% from CDN 2
There is no single best CDN in the same geographic
region or over time
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN Streaming Failures Are Common Events
% of stream failures: % of streams that failed to start
Three months dataset (May-July, 2011) for a premium
customer using Flash
CDN (relative) performance varies greatly over time
Opportunities for Improving Quality
Possible Actions to Improve Quality
Switch the bitrate
↓ Buffering, high frame drops, high start time, …
↑ High available bandwidth, …
Switch the CDN
↔ Connection error, missing content, buffering on low bitrate, ...
When to perform switching/selection?
Start time selection only
Start time selection & midstream switching
Potential Improvement Example: CDN Switching Only
For each CDN partition clients by
(ASN, DMA)
DMA: Designated Market Area
Potential Improvement Example: CDN Switching Only
For each CDN partition clients by
(ASN, DMA)
DMA: Designated Market Area
For each partition compute:
Buffering ratio
Failure ratio
Start time
….
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Potential Improvement Example: CDN Switching Only
For each CDN partition clients by
(ASN, DMA)
DMA: Designated Market Area
For each partition compute:
Buffering ratio
Failure ratio
Start time
….
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Avg. buff ratio of users in ASN[1]xDMA[1]
streaming from CDN1
Potential Improvement Example: CDN Switching Only
For each CDN partition clients by
(ASN, DMA)
DMA: Designated Market Area
For each partition compute:
Buffering ratio
Failure ratio
Start time
….
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Avg. buff ratio of users in ASN[1]xDMA[1]
streaming from CDN1
Avg. buff ratio of users in ASN[1]xDMA[1]
streaming from CDN2
Potential Improvement Example: CDN Switching Only
For each partition select best CDN
and assume all clients in the
partition selected that CDN
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Potential Improvement Example: CDN Switching Only
For each partition select best CDN
and assume all clients in the
partition selected that CDN
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
CDN1 >> CDN2
Potential Improvement Example: CDN Switching Only
For each partition select best CDN
and assume all clients in the
partition selected that CDN
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Potential Improvement Example: CDN Switching Only
For each partition select best CDN
and assume all clients in the
partition selected that CDN
Essentially, pick partition with best
quality across CDNs CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Best CDN (buffering ratio)
DMA
ASN
Potential Improvements
Provider1: large UGV (User Generated Video) site
Provider2: large premium VoD content provider
Base-line: existing assignment of viewers (clients) to CDNs
Metric Provider1 (UGV) Provider2 (Premium)
Base
line
Start-
time
Selectio
n
Mid-
stream
Switching
Base
line
Start-
time
Selection
Mid-
stream
Switching
Buffering
ratio (%)
6.8 2.5 1 1 0.3 0.1
Between x2.7 and x10 improvement in buffering ratio
Coordinated Control Plane for High
Quality Video Delivery
Video Control Plane Architecture
Coordinator implementing a global optimization algorithm that
dynamically select CDN & bitrate for each client based on
Individual client
Aggregate statistics
Content owner policies
(CDN/ISP info)
Content owners (CMS & Origin)
CDN 1
CDN 2
CDN 3
Clients
Coordinator
Co
nti
nu
ou
s m
easu
rem
ents
Bu
sin
ess
Polic
ies
control
Example: Local vs. Global Optimization
0
20
40
60
80
100
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000Ban
dw
idth
Flu
ctu
atio
n (
%)
Concurrent Viewers
Bandwidth fluctuation = (Max Bandwidth – Min Bandwidth)/(Average Bitrate)
Example: Local vs. Global Optimization
CDN1 DMA
ASN
DMA
ASN
DMA
ASN
CDN2
CDN3
0
20
40
60
80
100
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000Ban
dw
idth
Flu
ctu
atio
n (
%)
Concurrent Viewers
0
10
20
30
40
0 5000 10000 15000 20000 25000 30000 35000
Ban
dw
idth
Flu
ctu
atio
n (
%)
Concurrent Viewers
ASN/DMA saturated on all CDNs Don’t switch CDN; cap bitrates, instead
Concluding Remarks (I)
Key transition of main-stream video to the Internet
Video quality presents opportunity and challenge
Premium video on big screens zero tolerance for poor quality
Video player continuous monitoring and global optimization
has best chance of delivering high quality video
Many challenges remain, e.g.,
Scalability
How do multiple coordinators interact?
…
Concluding Remarks (II)
The video traffic dominance in the Internet is growing
Over 51% Internet traffic today, will be more than 86% in the next 4
years
The Internet is becoming a Video Network
Managing video delivery and maximizing video quality must
be at the core of any future Internet architecture!
Backup Slides
Conviva Optimization in the Wild
… increased average bit-rate from
1.7 Mbps to 2.1 Mbps…
Reduced views impacted by
buffering from 16.13% to 5.56%
…
… and raised engagement
by 36%
0,00%
5,00%
10,00%
15,00%
20,00%
25,00%
1600
1700
1800
1900
2000
2100
2200
Views Impacted by Buffering
Average Bit Rate
Possible Coordinator ArchitectureContinuous real-time measurements from every client
Inference Engine
Bit Rates
CD
Ns
Decision Engine
Optimize viewer performance by selecting the best option within the set of bit rates and CDNs
Akamai
DMA
ASN
DMA
ASN
DMA
ASN
Limelight
Level3 Time of day
Localize issues by region, network, CDN, and time
Real-time Global Data Aggregation and Correlation
Historical Data Aggregation and
Analysis
Global Inference, Decision & Policy
Engine
Real-time global optimizations
Conviva Services Enhance the Viewer Experience
and Lift Engagement by Lifting Bit Rate and
Reducing Buffering
Increased average bit-rate from 1.6 Mbps to 2.1 Mbps …
1.5%
1.0%
0.5%
DMS LAUNCH EMS LAUNCH
… reduced buffering ratio from 1.5% to 0.5%
… and raised engagement by 36%
Potential Improvements
Customer1: large UGV site
Customer2: large premium content provider
Note: * denotes improvements when using mid-stream
switching
Metric Customer1 Customer2
Current Projected Current Projected
Buffering ratio
(%)
6.8 2.5 / 1* 1 0.3 / 0.1*
Start time (s) 6.41 2.91 1.36 0.9
Failure ratio (%) 16.57 2.4 1.1 0.7
Between x2.7 and x10 improvement in buffering ratio
Video Quality Matters [Sigcomm’11]
Quality has substantial impact on viewer engagement
Need to ensure uninterrupted streaming at high bitrates
Buffering ratio is most critical across genres
Highest impact for live: 1% of buffering reduced play time by 3min
1% increase in buffering leads to more than 60% loss in
audience
1% difference in buffering between two ISPs
68% monthly loss in uniques for ISPwith poor performance
Customer1: Start-time vs. Midstream CDN
Switching
78%
84%
90%
Provider1: Oracle vs. Historical
Base-lineOracle
Historic
Potential Improvement Example: CDN Switching Only
Oracle:
For each partition select best CDN
and assume all clients in the partition
selected that CDN
Essentially, pick partition with best
quality across CDNsCDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
Best CDN (buffering ratio)
DMA
ASN
Potential Improvement Example: CDN Switching Only
Details
If a partition has not enough clients
use a larger partition
?
CDN1 (buffering ratio)
DMA
ASN
Potential Improvement Example: CDN Switching Only
Details
If a partition has not enough clients
use a larger partition
Use quality metric distribution to
predict quality of a client on new
CDNCDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN
CDNs Vary in Performance over Geographies and Time
0,0%
10,0%
20,0%
30,0%
40,0%
50,0%
60,0%
70,0%
80,0%
90,0%
100,0%
Was
hin
gto
n D
C (
Hag
erst
ow
n):
CM
CS(
33
65
7)
Milw
auke
e:R
OA
DR
UN
NER
-CEN
TRA
L(2
02
31
)
Gre
en B
ay -
Ap
ple
ton
:SC
RR
-70
15
(70
17
)
Den
ver:
ASN
-QW
EST(
20
9)
Ch
arlo
tte:
SCR
R-1
14
26
(11
42
6)
Was
hin
gto
n D
C (
Hag
erst
ow
n):
ASN
-CX
A-A
LL-C
CI-
22
773
-…
Ph
ilad
elp
hia
:VZG
NI-
TRA
NSI
T(1
92
62
)
San
Die
go:S
BIS
-AS(
71
32
)
Las
Veg
as:A
SN-C
XA
-LV
-13
43
2-C
BS(
13
43
2)
Mad
iso
n:C
HA
RTE
R-N
ET-H
KY-
NC
(20
115
)
Ind
ian
apo
lis:S
BIS
-AS(
71
32
)
Pro
vid
ence
-N
ew B
edfo
rd:A
SN-C
XA
-ALL
-CC
I-2
277
3-…
Was
hin
gto
n D
C (
Hag
erst
ow
n):
VZG
NI-
TRA
NSI
T(1
92
62
)
Har
tfo
rd &
New
Hav
en:S
BIS
-AS(
71
32
)
Ho
ust
on
:SB
IS-A
S(7
13
2)
Gra
nd
Rap
ids
-K
alam
azo
o -
Bat
tle
Cre
ek:S
BIS
-AS(
71
32
)
Atl
anta
:BEL
LSO
UTH
-NET
-BLK
(638
9)
Ho
no
lulu
:HA
WA
IIA
N-T
ELC
OM
(36
149
)
Atl
anta
:CO
MC
AST
-77
25
(77
25)
Was
hin
gto
n D
C (
Hag
erst
ow
n):
SPC
S(1
05
07
)
Orl
and
o -
Day
ton
a B
each
-M
elb
ou
rne:
EMB
AR
Q-…
Den
ver:
CM
CS(
33
65
2)
No
rfo
lk -
Po
rtsm
ou
th -
New
po
rt N
ews:
ASN
-CX
A-A
LL-C
CI-
…
Sain
t Lo
uis
:CH
AR
TER
-NET
-HK
Y-N
C(2
01
15
)
Cin
cin
nat
i:FU
SE-N
ET(6
18
1)
Ph
oen
ix:A
SN-Q
WES
T(2
09
)
Dal
las
-Fo
rt W
ort
h:V
ZGN
I-TR
AN
SIT(
19
26
2)
Pit
tsb
urg
h:C
CC
H-3
(70
16
)
Sain
t Lo
uis
:SB
IS-A
S(7
13
2)
San
Fra
nci
sco
-O
akla
nd
-Sa
n J
ose
:SB
IS-A
S(7
13
2)
San
Die
go:R
OA
DR
UN
NER
-WES
T(2
00
01
)
Gre
envi
lle -
Spar
tan
sbu
rg -
Ash
evill
e -…
Kan
sas
Cit
y:SB
IS-A
S(7
13
2)
Co
lum
bu
s -
OH
:SC
RR
-10
79
6(1
07
96
)
Lou
isvi
lle:I
NSI
GH
T-C
OM
MU
NIC
ATI
ON
S-C
OR
P-A
S1(3
67
27)
Ch
icag
o:A
TT-I
NTE
RN
ET3(
64
78
)
Kan
sas
Cit
y:A
SN-C
XA
-ALL
-CC
I-2
277
3-R
DC
(22
77
3)
Mia
mi -
Fort
Lau
der
dal
e:B
ELLS
OU
TH-N
ET-B
LK(6
38
9)
Cle
vela
nd
:NEO
-RR
-CO
M(1
10
60
)
Ch
icag
o:V
ZGN
I-TR
AN
SIT(
19
262
)
Co
lum
bia
-SC
:SC
RR
-11
42
6(1
14
26
)
Dal
las
-Fo
rt W
ort
h:S
BIS
-AS(
71
32
)
Det
roit
:SB
IS-A
S(7
13
2)
Kan
sas
Cit
y:SC
RR
-11
95
5(1
19
55
)
Los
An
gele
s:C
HA
RTE
R-N
ET-H
KY-
NC
(201
15
)
Mia
mi -
Fort
Lau
der
dal
e:C
OM
CA
ST-2
02
14
(20
21
4)
Cle
vela
nd
:SC
RR
-10
79
6(1
07
96
)
Wes
t P
alm
Bea
ch -
Fort
Pie
rce:
CO
MC
AST
-20
21
4(2
02
14)
San
Die
go:A
SN-C
XA
-ALL
-CC
I-22
77
3-R
DC
(22
773
)
Seat
tle
-Ta
com
a:A
SN-Q
WES
T(2
09)
New
Yo
rk:C
AB
LE-N
ET-1
(61
28
)
Po
rtla
nd
-O
R:A
SN-Q
WES
T(2
09
)
Okl
aho
ma
Cit
y:SB
IS-A
S(7
13
2)
Ph
oen
ix:A
SN-C
XA
-ALL
-CC
I-2
277
3-R
DC
(22
77
3)
Ch
icag
o:S
BIS
-AS(
71
32
)
Gre
ensb
oro
-H
igh
Po
int
-W
inst
on
-Sal
em:S
CR
R-…
Alb
any
-Sc
hen
ecta
dy
-Tr
oy:
RR
-NYS
REG
ION
-ASN
-01
(11
35
1)
Tyle
r -
Lon
gvie
w (
Lufk
in &
Nac
ogd
och
es):
SUD
DEN
LIN
K-…
Gra
nd
Rap
ids
-K
alam
azo
o -
Bat
tle
Cre
ek:C
MC
S(3
36
68
)
Ho
ust
on
:CM
CS(
33
66
2)
Har
tfo
rd &
New
Hav
en:C
OM
CA
ST-7
01
5(7
01
5)
Euge
ne:
ASN
-QW
EST(
209
)
San
Fra
nci
sco
-O
akla
nd
-Sa
n J
ose
:CM
CS(
33
65
1)
Potential Improvement Example: CDN Switching Only
Oracle:
For each partition select best CDN
and assume all clients in the partition
selected that CDN
Historical:
For each partition select best CDN in
previous epoch, and assign clients to
that CDN in next epoch
CDN1 (buffering ratio)
DMA
ASN
CDN2 (buffering ratio)
DMA
ASN