VOD: Video On Demand
The materials presented here are adapted
from Dave Sincoskie's pioneer paper, "System Architecture for a Large Scale
Video On Demand Service," Computer Network and ISDN Systems, 22, (1991)
155-162. The topic is related to real-time, multimedia storage/communications
Video On Demand Example
Video On Demand Service
Provides service similar to the popular
video-tape rental service.
The term, Interactive
video, covers broader services.
It is provided from a centralized location
over a digital network.
A customer can signal the network to select
a movie, the network will start transmission within a reasonable time.
Coarse rewind and fast forward can also
All customers could choose to watch the
same movie but individual customers may be watching different part of the
The North American telephone systems have
a common 1.544 Mbps rate.
It is possible to extend this rate into
the subscriber loop, which is the last leg of the connection to the customer.
For broadband, the lowest channel rate
of SONET standard is 51.84 Mbps.
Assume two coding rates:
NTSC video encoded into a 1.5 Mbps channel.
HDTV encoded into a 50 Mbps channel.
NTSC service over upgraded copper technology.
HDTV service over BISDN fiber loop.
Use double star topology: with non-shared
fiber running from the customer premise to a remote concentrator, and then
shared fibers from the concentrator to the local central office.
The central office switch is an ATM-based
packet switch with multicast capability. The centralized video database
is connected to the central office with a large amount of dedicated fiber
The video on demand service may be possible
sometime before year 2000. (This was predicted in 1991, -- it was pushed
hard during 94-96 in several field trials but pace slows down due to network
architecture cost/price problems.)
Direct TV and Prime Star are distributing
video using MPEG2.
consists of a large number of movies with
a length about 100 minutes each.
1.125/37.5 GBytes for 100 minutes at 1.5/50
(100*60*1.5/8=1.125 GB; 100*60*50/8=37.5GB)
video may be stored on video tapes or video
D3, SVHS VTR tape are used in TCI.
video may in analog or digital form but is
converted into a digital stream at a constant rate of 1.5/50 Mbps. (mpeg
stream has variable bit rate)
the customer is responsible for conversion
from digital to display on his TV.
Three-level memory hierarchy will be used
to store the video database.
Disk Technology Trend
1998 35 GB based on annual increase
1998 AV Disk 9GB common
The transfer rate is controlled by the
linear bit density and rotational speed.
Ultra wide SCSI drives transfer data at
40MBps (8ms access time, 512k cache, 9.1GB, $800).
Enhanced IDE and Ultra DMA drives transfer
data at 33MBps (9.5ms access time, 512k cache, 6.4GB, $259)
RAID data striping can multiply the transfer
speed of single drive.
The real-time Audio/Video applications
require fast response time and sustain delivering speed. Disk cache at
hard drive may have to be turned off or speeded up to meet the requirements.
Assume no change in rotational speed,
both linear density and radial density increase twice resulting four times
increase in disk density,
by 1998 the disk may have 83 Mbps transfer
Central Office Switch
For 10,000 customers each watches 1.5/50 Mbps
video stream, we need 15/500 Gbps.
The switch design based on Batcher-banyan
technology and 3D packaging is capable of delivering 2.4 Tb/s with 16K
Video on Demand System Architecture
Library is read-on memory.
Each copier has one single input (from
tape player) and one output per disk head.
Each stop-start buffer has one input and
Inputs and outputs are multiplexed up to
the switch rate of 150 Mbps or higher
View with More Detailed Multimedia Storage
A single customer watching a movie that
no other customer is viewing. A connection is setup between customer and
A popular movie, when the first customer
requests it, it will be cached into the copier memory.
The network allocates a copier disk to
signals the library to load the tape,
set up a multicast connection from the
library tape to the copier disk and also to the customer.
sometime later, customer 2 requests the
customer 2 may needs to wait until disk
heads move back to their beginning position. (For example, customer 1 uses
4th disk head while customer 2 uses the 1st head. Assume a disk may have
20 heads, each head covers 5 min of the movie).
If customer 3 requests the same movie when
customer 2 is waiting, the network can set up a multicast connection to
connect customer 2 and 3 to the same head.
This brings the customers into a fixed
phase relationship, with granularity controlled by the number of heads
per movie. Each phase contains 5 min movie (100min/20 heads=5min).
For HDTV case, each disk holds a movie
(37.5 GB) and has 20 heads operating in parallel at 50 Mbps each. The relative
position of the heads are fixed.
Each head hold 5min movie.
The head phase difference is 5 min.
For NTSC case, assume
We use two 637MB disks, 4 head each.
2*637MB=1.274GB > 1.125 GB for 100
min NTSC movie.
Each head has a transfer rate of 20.4Mbps.
With 4 heads, the disk can transfer 20.4*4=81.6Mbps.
Each head can produce max. 13 phases at
1.5 Mbps per phase.
With 8 heads per movie, 13*8=104 different
phases is achieved.
Phase difference=100*60/104=57.7 sec/phase.
To make the frame number dividable by 8/13,
add 24 frame to make it 180024.
Each head has 180024/8=22,503 frame and
grouped into 22503/13=1731 blocks. each block has 13 frames and each frame
holds 1 frame from each of the 13 phases.
Assume uniformed frame size. Each frame
How long for a disk head to transfer the
12.5min length of the movie that it covers?
12.5*60s*1.5Mbps/20.4Mbps=55.14 s or
1731*13*50kb/20.4Mbps = 1125.15Mb/20.4Mbps
= 55.15 s.
Is it correct that since the disk head
covers 12.5 min movies, it will take 12.5min to read the same frame again?
A good research project will be to modify
the above design to allow full function rewind and fast forward.
It allows a customer to pause and resume.
The size of the stop-start buffer is twice
the phase difference (57.7*2=115.4)/(5min*2=10*60=600) s or (1.5Mbps*115.4/8=21.6MB)/3750
MB. The bandwidth is twice the coding rate at 3/100 Mbps.
In HDTV case, it contains 1/10 of the
In NTSC case, it contain 2/104 of the
movie 100*60*2/104 = 115.4s
When customer 2 requests a pause, the network
allocates a stop-start buffer.
redirect the transmission from the disk
to this buffer. (change the connection)
When customer 2 requests resume, the data
comes from the head of the buffer while the transmission from the disk
continues into the tail end.
Pauses longer than twice the head phase
difference can be handled by a combination of the stop-start buffer and
a head change.
It provides limited visual rewind or fast
forward. (There may be problems with this design.)
The stop-start buffer must be allocated
on a per-user basis.
Therefore it may be placed in the customerís
TV or digital decoder.
Bonus Exercise: Discuss the impact
of including fast rewind and fast forward on the memory requirement of
VoD and propose the design of a stop-start buffer and its operation for
handling such new features.
The above discussion is only qualitative.
Quantitative studies will require considerably
more information on cost and performance of various components, and the
statistics of customerís viewing habits.
Considering the placement of a video-on-demand
system in a LATA (Local Access and Transport Area) instead of class 5 local
Assume 625,000 customer homes per LATA, 60
class 5 offices per LATA, 2000 movies available.
eliminate the library and stop-start buffers,
keep movies in disks.
$10,000 cost per disk
The cost of the video database would come
out to ($10,000*2000/625000=)$32 per customer, only small fraction of fiber
loop and loop electronics.
Assume every phase of every movie is multicast
to every class 5 office, this would require a maximum of 324/2000 Gbps
For NTSC, 2000movies*104phases/movie*1.5Mbps/phase=312Gbps.
For HDTV, 2000movies*20phases/movie*50Mbps/phase=2000Gbps.
The number of channels is much bigger than
that of existing broadcast TV channel.
NTSC video-on-demand system is well within