A Server-less Architecture for Building Scalable, Reliable, and Cost-Effective Video-on-demand Systems

1

• A Server-less Architecture for Building Scalable,
  Reliable, and Cost-Effective Video-on-demand
  Systems
• Raymond Leung and Jack Y.B. Lee
• Department of Information Engineering
• The Chinese University of Hong Kong

2
Contents

• Introduction
• Server-less Architecture
• Performance Evaluation
• System Scalability
• Summary

3
Client-Server Architecture
Introduction

• Traditional client-server architecture
• clients connect to server for streaming
• system capacity limited by server capacity

4
Motivation
Introduction

• Limitation of client-server system
• system capacity limited by server capacity
• high-capacity server is very expensive
• Availability of powerful client-side device, or
  called set-top box (STB)
• home entertainment center - VCD/DVD player,
  digital music jukebox, etc.
• relatively high processing capability, and local
  HD storage
• Server-less architecture
• eliminates the dedicated server
• each user node (STB) serves both as a client and
  as a mini-server
• fully distributed storage, processing, and
  streaming

5
Server-less Architecture
Architecture

• Basic principles
• dedicated server is eliminated
• users are divided into clusters
• video data is distributed to nodes in a cluster

6
Challenges
Architecture

• Data placement policy
• Retrieval and transmission scheduling
• Fault tolerance
• Distributed directory service
• System adaptation and dynamic reconfiguration
• etc.

7
Data Placement Policy
Architecture

• Block-based striping
• video data is divided into fixed-size blocks and
  then distributed among nodes in the cluster
• low storage requirement, load balanced
• capable of fault tolerance using redundant
  unit(s)

8
Retrieval and Transmission Scheduling
Architecture

• Round-based Schedulers
• retrieves data block in each micro-round
• transmission starts at the end of micro-round

9
Retrieval and Transmission Scheduling
Architecture

• Disk retrieval scheduling
• Grouped Sweeping Scheme1 (GSS)
• able to control the tradeoff between disk
  efficiency and buffer requirement
• Transmission scheduling
• Macro round length
• time required that every node sends out a data
  block of Q bytes
• depends on system scale, data block size and
  video bitrate

Tf macro round length n number of nodes
within a cluster Q data block size Rv video
bit-rate
1P.S. Yu, M.S. Chen D.D. Kandlur, Grouped
Sweeping Scheduling for DASD-based Multimedia
Storage Management, ACM Multimedia Systems, vol.
1, pp. 99 109, 1993
10
Retrieval and Transmission Scheduling
Architecture

• Transmission scheduling
• Micro round length
• under the GSS scheduling, the GSS group duration
  within each macro round
• depends on macro round length and number of GSS
  groups

Tg micro round length Tf macro round length g
number of GSS groups
11
Fault Tolerance
Architecture

• Node characteristics
• lower reliability than high-end server
• shorter mean time to failure (MTTF)
• system fails if any one of the nodes fails
• Fault tolerance mechanism
• erasure correction code to implement fault
  tolerance
• Reed-Solomon Erasure code2 (RSE)
• retrieve and transmit coded data at higher data
  rate
• recover data blocks at the receiver node

2A. J. McAuley, Reliable Broadband Communication
Using a Burst Erasure Correcting Code, in Proc.
ACM SIGCOMM 90, Philadelphia, PA, September 1990,
pp. 287306.
12
Fault Tolerance
Architecture

• Redundancy
• encode redundant data from video data
• recover lost data in case of node failure(s)

13
Performance Evaluation
Performance Evaluation

• Storage capacity
• Network capacity
• Disk access bandwidth
• Buffer requirement
• System response time

14
Storage Capacity
Performance Evaluation

• What is the minimum number of nodes required to
  store a given amount of video data?
• For example
• video bitrate 150 KB/s
• video length 2 hours
• storage required for 100 videos 102.9GB
• If each node can allocate 1GB for video storage,
  then
• 103 nodes are needed (without redundancy) or
• 108 nodes are needed (with 5 nodes added for
  redundancy)
• This sets the lower limit on the cluster size.

15
Network Capacity
Performance Evaluation

• How many nodes can be connected given a certain
  network switching capacity?
• For example
• video bitrate 150KB/s
• If the network switching capacity is 32Gbps, and
  assume 60 utilization
• up to 8388 nodes (without redundancy)
• Network switching capacity is not a bottleneck.

16
Disk Access Bandwidth
Performance Evaluation

• Recall the retrieval and transmission scheduling
• Continuous data transmission constraint
• must finish retrieval before transmission in each
  micro-round
• need to quantify the disk retrieval round length
  and verify against the above constraint

17
Disk Access Bandwidth
Performance Evaluation

• Disk retrieval round length
• time required retrieving data blocks for
  transmission
• depends on seeking overhead, rotational latency
  and data block size
• suppose k requests per GSS group
• Continuous data transmission constraint

18
Disk Access Bandwidth
Performance Evaluation

• Example
• Disk Quantum Atlas 10K3
• Data block size (Q) 4KB
• Video bitrate (Rv) 150KB/s
• Number of nodes N
• GSS group number (g) N (reduced to FCFS
  scheduling)
• Micro round length
• Disk retrieval round length 0.017s lt 0.027s
• Therefore the constraint is satisfied even if
  FCFS scheduler is used.

3G. Ganger and J. Schindler, Database of
Validated Disk Parameters for DiskSim,
http//www.ece.cmu.edu/ganger/disksim/diskspecs.h
tml
19
Buffer Requirement
Performance Evaluation

• Receiver buffer requirement
• double-buffering scheme
• one for storing data received from the network
  plus locally retrieved data blocks
• another one for video decoder
• Sender buffer requirement
• under GSS scheduling

20
Buffer Requirement
Performance Evaluation

• Total buffer requirement versus system scale
• Data block size 4KB, Number of GSS groups gN

21
System Response Time
Performance Evaluation

• System response time
• time required from sending out request to
  playback begins
• scheduling delay pre-fetch delay
• Scheduling delay under GSS
• time required from sending out request to data
  retrieval starts
• can be analyzed using urns model
• detailed derivation available elsewhere4
• Prefetch delay
• time required from retrieving data to playback
  begins
• one micro round to retrieve a data block and one
  macro round to transmit the whole block to the
  client node

4Lee, J.Y.B., Concurrent push-A scheduling
algorithm for push-based parallel video servers,
IEEE Transactions on Circuits and Systems for
Video Technology, Volume 9 Issue 3 , April
1999, Page(s) 467 -477
22
System Response Time
Performance Evaluation

• For example
• Data block size 4KB

23
System Scalability
System Scalability

• Not limited by network or disk bandwidth
• prefers FCFS disk scheduler over SCAN
• Limited by system response time
• prefetch delay increases linearly with system
  scale
• example response time of 5.615s at a scale of
  200 nodes
• Solution
• forms new clusters to expand system scale
• uses smaller block size (limited by disk
  efficiency)

24
Summary
Summary

• Server-less architecture proposed for VoD
• dedicated server is eliminated
• each node serves as both a client and a
  mini-server
• inherently scalable
• Challenges addressed
• data placement policy
• retrieval and transmission scheduling
• fault tolerance
• Performance evaluation
• acceptable storage and buffer requirement
• scalability limited by system response time

25

• End of Presentation
• Thank you
• Question Answer Session

26
Reliability
Appendix

• Higher reliability achieved by redundancy
• each node has independent failure and recovery
  rate, and respectively
• let state i be the system state where i out of
  the N nodes failed
• at state i, the changing rate to state (i1) and
  (i-1) are and respectively
• assume the system can tolerate up to h failures
  using redundancy
• the system state diagram is shown as follows

27
Reliability
Appendix

• System mean time to failure (MTTF)
• can be analyzed by continuous time Markov Chain
  model
• solving the expected time from state 0 to state
  (h1) in previous diagram,

28
Impact of Redundancy
Appendix

• Bandwidth requirement (without redundancy)
• (N-1) received from network and one locally
  retrieved from disk
• Bandwidth requirement (with h redundancy)
• additional network bandwidth will be needed for
  transmitting the redundant data

Rv video bit-rate
29
Impact of Redundancy
Appendix

• Data block size (without redundancy)
• block size Q bytes
• Data block size (with h redundancy)
• block size