Noriaki Kamiyama, Ryoichi Kawahara, Tatsuya Mori, Haruhisa

Multicast Pre-Distribution in
VoD Services
Noriaki Kamiyama,
Ryoichi Kawahara,
Tatsuya Mori,
Haruhisa Hasegawa
NTT Service Integration Laboratories
2011. 5. 11
Problem in VoD System
Demand in VoD concentrates at night and
changes widely on daily basis.
Maximum active session
count in each hour [102]



Max. active session count
in each hour for seven days
from Dec. 1, 2004 in VoD
service of China Telecom
25
20
15
10
5
0
12/1 2
3
4 5
Date
6
7
Demand at peak hours is
five times larger than that at
off-peak hours.
Need to design VoD system based on demand at peak
hours to maintain stable service quality
Important for ISPs to reduce server and network load at
peak hours to suppress network cost
2
Multicast Pre-Distribution (MPD) VoD


Users are provided huge amount of access link
bandwidth much larger than playback rate.
Large part of downlink bandwidth remains unused.

Effective to always deliver content and store them at
STBs independently of user requests

Video server can freely select delivery content and
freely set time at which each content is delivered.
 Delivers some popular content to all users in
multicast session

Propose MPD VoD system
3
Overview of MPD VoD




Video server always pre-distributes popular content to
all users, and STBs store all of them.
Determines pre-distribution schedule on daily basis
Users can watch pre-distributed content without
generating load at server and network at request.
Delivers content on-demand in unicast session for
requests of content not pre-distributed
On-demand
unicast delivery
TV
Network
STB
Server
Multicast pre-distribution
Watching pre-distributed content
4
Assumptions

Set transmission rate of each MPD channel as
Rb = Ad – R (Ad: downlink capacity, R: playback rate)

Video server makes pre-distribution schedule for 24
hours in n-th day at beginning of n-th day (0:00)

To preserve copyright of content, remove pre-distributed
content from STBs at 0:00 after L days elapse
5
Selecting Pre-distributed Content

To improve reduction effect of server and network load in
n-th day, desirable to pre-distribute content m with large
Dm,n, total download time (TDLT) for content m in n-th day

Estimates Dm,n using EMA (exponential moving average):

Dm,n = aDm,n-1 + (1 - a)Dm,n-1
(a: smoothing parameter)

Except content pre-distributed between (n-L)-th day and
(n-1)-th day, selects content in descending order of Dm,n

Pre-distributes selected content
in descending order of

estimated request count Dm,n/Sm (Sm: length of content m)
6
Access Log Data Used in Evaluation




Access log data in PowerInfo VoD system, commercial VoD
service of China Telecom
Seven months from 2004/6/1 to 2004/12/31
Request count: 20,921,657, content count: 6,735
User count: 37,360, average content size: 3,510 seconds
1
1
10-1
10-2
10-3
10-4
1
CCD
CCD
10-1
40% content was
45 minutes.
Another 40%
content was 90
minutes.
10
102
Content size (minutes)
Content
10-2
10-3
10-4
103
Users
Few users generated
many requests. Many
requests concentrated
on specific content.
10-5 -5 -4 -3 -2 -1
10 10 10 10 10 1 10 102 103
Average TDLT per day (hours)
7
Evaluation Condition





Transmission capacity of downlink: Ad = 10 Mbps
Playback rate: R = 2 Mbps
Transmission rate of MPD channel:
Rb = Ad – R = 8 Mbps
Average content count pre-distributed on each day:
X = 98.46
Smoothing parameter of EMA: a = 2/3
(Data of most recent two days occupied 86% of
information used in estimation.)
8


25
L=0
20
15
10
L=1
5
L=7
0
6/1 7/1 8/1 9/1 10/1 11/1 12/1 1/1
Date
Maximum active session
count in each hour [102]
Maximum active session
count in each day [102]
Effect of MPD on Server Load
25
L=0
L=1
20
15
10
5
0
12/1
2
L=7
3
4
5
Date
6
7
Max. session count was largely reduced by MPD
compared with only unicast delivery (L = 0).
Effect of MPD was stronger when demand was larger.
9
Effect of MPD on Network Load
(Av. traffic on each link using MPD) / (That using only unicast delivery)
Relative average
amount of traffic
1.0

0.8

0.6
0.4
Server load
0.2
0
0
1
2
3
4
5

6
Used 31 topologies of
commercial ISPs
Allocated server at node
giving minimum av. hop
length of unicast flows
Transmitted packets on
minimum spanning tree
7
L: content life length in STBs [days]


Average link load was reduced by MPD in all 31 networks.
The effect of MPD was larger as L increased.
10
Required Storage Size of STBs
L
Kmax
V(107)
(%)
Zmax (TB)
0
2409
8.79
***
***
1
1738
6.34
0.231
0.69
2
1488
5.26
0.331
1.38
3
1312
4.62
0.389
2.07
4
1174
4.15
0.432
2.76
5
1059
3.77
0.465
3.46
6
972
3.45
0.492
4.15
7
916
3.18
0.514
4.84

As L increased, Kmax and V decreased.
However, as L increased, Zmax linearly increased.
Just small part of pre-distributed content were viewed.






Kmax: max. active session
count over entire log period
V: average total session
length in each day
: utilization of
pre-distributed content
Zmax: max. required storage
size of STBs
11
Conclusion

Important to reduce peak load in VoD service

Proposed to pre-distribute popular content to all users
on multicast session independently of actual user
requests, in addition to on-demand unicast delivery

Numerical evaluation using access log of actual VoD:
 Reduced server and network load
 Reduction effect of peak server load:
30% (L = 1), 60% (L = 7)
 Needed several TB storage at STBs
12
Merit of MPD VoD

Reduction of load of video server and network



On-demand service & simple VCR operations



Pre-distributes popular content
 Reduces number of on-demand deliveries
Multicast pre-distribution
 Suppresses load cause by pre-distribution
Users view pre-distributed content from STBs.
On-demand delivery for content not pre-distributed
No need for user cooperation

Only user owning STB can view pre-distributed content
13
VoD Service

Increase of transmission capacity of access links
 Dramatically increasing user count of VoD services
Request
Content delivery
Server
Network
TV

VoD delivering UGC (user generated content):


STB(set-top box)
Provided by service providers independent of ISPs
VoD delivering rich content:

Provided mainly by ISPs regarding it as important service
14
CDN


Provides multiple cache servers over network and
distributes copies of content on cache servers
Delivers content from cache server close to
requesting user
TV
Network
Original
server
STB
Cache server

Total cost of server facilities is not reduced because
total load over all server facilities is constant.
15
Multicast Delivery System

Delvers content to multiple users requesting same
content on single multicast delivery session
Network
TV
Server
STB


Increases user response time to increase user count
accommodated in each multicast session
Needs complex operations to switch delivery session
to support VCR operations
16
P2P Delivery System

Each peer, i.e., user, uploads viewing content to other
peers viewing same content
Network
TV
Server
STB


Needs to sustain content delivery when uploading
user leaves or making VCR operations  Needs
complex operations, i.e., switching source peers
Needs incentive for users
17
サーバ負荷低減法(4): ISP制御型P2P配信



HDレコーダーの価格低下は目覚ましく,現在,1Tバイトの容量のHD
レコーダーを4万円程度で購入可能
ユーザ宅内のSTBをISPの管理下で大容量のキャッシュとして運用
ユーザ離脱への対応が不要で全体最適化が可能
Network
TV
Server
STB


ユーザに対してISP管理下のSTBを設置させ,他のユーザへの配信
に要する電力を負担するインセンティブの付与が必要
配信サーバの負荷は低減されるが,NWの負荷は低減せず
18
総視聴時間上位コンテンツの一致度
d日間離れた日の各総視聴時間上位x個のコンテンツの中で,両日と
もに上位x以内となったコンテンツの割合
Av. duplication ratio of top x
content between lag d days


1
d=1
d=2
0.8 d = 3
0.6
d=4
d=5
0.4
d=6
0.2
d=7
0
1
10
102
x
103
104
xが小さい領域では,間隔dが増加するに伴い総視聴時間上位x個の
コンテンツの一致度は低下するが,xの増加に伴い一致度は増加し,
xが1,000程度以上になるとdによる一致度の差異は消滅
19
BC事前配信コンテンツの選択効率
各日にx個をBC事前配信した際の選択効率:
(STB上のLx個の総視聴時間) / (実際の上位Lx個の総視聴時間)
Average efficiency of
selecting top x content

1
0.8
0.6
0.4
L=1
L=3
L=7
0.2
0
1
10
102
x
103
104
xが小さい場合,Lの増加に伴い,BC事前配信時点と各日の総視聴時
間上位コンテンツの一致度が低下し,選択効率は低下

しかし選択効率の低下度合いは小さく,Lが7日でも70%の選択効率

xの増加に伴い選択効率は向上し,x  10であれば85~90%を達成
⇒ EMAを用いた総視聴時間の推定値に基づきBC事前配信コンテンツを
選択することで,十分な選択効率が達成可能
20
