Optimizing BitTorrent for OnDemand Streaming

1
Optimizing BitTorrentfor On-Demand Streaming

• Nadim Parvez (U. Calgary)
• Carey Williamson (U. Calgary)
• Anirban Mahanti (IIT Delhi)
• Niklas Carlsson (U. Saskatchewan)

(To appear at ACM SIGMETRICS 2008, Annapolis, MD,
June 2008.)
2
Introduction and Motivation

• BitTorrent is a popular protocol used for
  Peer-to-Peer (P2P) file sharing on the Internet
• Very efficient for download of large files, such
  as media objects (audio/video)
• Question How suitable is BitTorrent for
  on-demand stored media streaming?

3
Background BitTorrent 101

• Key features of BitTorrent
• a single file (e.g., 100 MB movie) is split
  into many fixed-size pieces (e.g., 256 KB)
• all peers associated with a file are called
  a swarm (tracker, downloaders, and seeds)
• peers can download pieces in any order
• Rarest-First piece selection policy makes
  prevalence of pieces asymptotically uniform
• can download pieces concurrently from one or
  more peers (typically 4-5 at a time)
• tit-for-tat reciprocity among peers to
  encourage cooperation (upload/download)

4
Background Streaming

• Download Streaming
• D obtain file first, then do something with it
• S view the file while it is still being obtained
  (start-up delay
• Media playback must be sequential
• Media file has an inherent media playback rate
  (frames per second)
• Playback rate must be sustained for the duration
  of the media object (uninterrupted)

5
Background P2P Streaming

• The P2P paradigm has also been applied to media
  streaming applications
• CoolStreaming, P2PLive, ZigZag,
• Two types of streaming
• live streaming usually a single source, and many
  peers with shared temporal content focus, joining
  and leaving at any time
• on-demand streaming stored media objects,
  accessed at any time, retrieved in entirety
• Our focus on-demand P2P streaming

6
Research Questions

• Can BitTorrent-like protocols provide scalable
  on-demand streaming?
• How sensitive is the performance to the
  application configuration parameters?
• Piece selection policy
• Upload/download bandwidth
• What is the user-perceived performance?
• Startup delay
• Probability of disrupted playback

7
Some Theory Highlights

• Mathematical modeling/analysis of BitTorrent-like
  P2P system
• Fluid flow modeling of system behaviour
• Differential equations describing evolution of
  system population (transient and/or
  steady-state analysis)
• Combinatorics and optimization
• Possible game-theoretic analysis
• Possible security/vulnerability analysis

8
A Key Observation

• MediaStreamingProgress (MSP) depends on two
  different things
• DownloadProgress (DP)
• SequentialProgress (SP)
• These two can be analyzed separately

MSP DP x SP
(useful media pieces per unit time)
(pieces obtained per unit time)
(useful media pieces per pieces obtained)
9
Quiz Time Sequential Progress
Q After having retrieved k pieces (at
random) from a file with M pieces, what is the
probability that a peer has (exactly) pieces 1
to j inclusive?
A1 Prob(M,k,j)
A2 P
10
Sequential Progress Example
11
BitTorrent Model (1 of 2)
Downloader
Seed
Downloader
Seed
Downloader
Torrent (with x downloaders and y seeds)
Arrival rate ?
Departure rate ? y
Downloader
Residence Time 1/ ?
Download Time T
12
BitTorrent Model (2 of 2)
Seed
Downloader
Seed
Downloader
U of upload connections C Per connection
throughput
Downloader
Torrent (with x downloaders and y seeds)
Arrival rate ?
Departure rate ? y
D of download connections C Per connection
throughput
Downloader
13
Assumptions and Parameters

• Single swarm homogeneous peers
• x downloaders and y seeds at time t
• D download conns U upload conns
• System is demand-driven xD (xy)U
• Download latency T
• Number of pieces in the file M
• Startup delay ?
• Media Playback Rate r

14
Fluid Model Overview
Arrivals
x(t)
Conversions
Downloaders
y(t)
Departures
Seeds
15
Model Rarest-First

• Conversion of downloaders to seeds at rate
  (xy)UC.
• Therefore the change of swarm population

16
Model Rarest-First

• Download latency
• Sequential progress
• Startup delay

17
Rarest-First Observations

• Download latency is independent of the peer
  arrival rate (system is scalable).
• Latency improves when upload bandwidth or seed
  residence time increase.
• Sequential progress is very poor (bad
  news for on-demand streaming!).
• Startup delay approximately equals the download
  latency when M is large.

18
Strict In-Order Model (naïve)

• Obtain pieces in numerical order
• Best case for sequential progress
• Implications
• Completely breaks the tit-for-tat mechanism!
  (only older peers have useful pieces
  for you)
• Uneven distribution of demand in system (heaviest
  at seeds and older peers, who use random
  selection with purging to handle load)
• Young peers progress quickly, but progress gets
  slower and slower with age
• Unstable system population (oscillation)

19
BitTorrent Model (revisited)
Seed
Downloader
Seed
Downloader
U of upload connections C Per connection
throughput
Downloader
Torrent (with x downloaders and y seeds)
Arrival rate ?
Departure rate ? y
Peers (sorted by age)
20
Model Strict In-Order

• The probability of getting a download connection
  for a downloader of age t is
• Averaging over all downloaders (age 0 to T),
  the change of swarm population is

21
Model Strict In-Order

• Swarm population
• Download latency
• Startup delay

22
Strict In-Order Observations

• System progress is very sluggish due to x/2 term
  in the conversion rate.
• Number of downloaders is much higher.
• Number of seeds is same as Rarest-First.
• Download latency is almost double compared to
  Rarest-First (for ? near 1).

23
In-Order(FCFS) Model

• Obtain pieces in numerical order
• Best case for sequential progress
• Features
• Pending requests are queued
• Queues are served in FCFS order
• Closed-loop mechanism for new requests ensures
  finite queue size
• Load is still uneven, but self-regulated
• System is fair to peers of all ages

24
Model Strict In-Order (FCFS)

• Swarm population
• Download latency
• Startup delay

25
In-Order(FCFS) Observations

• Population evolution is identical to
  Rarest-First.
• Download latency is identical to Rarest-First.
• Sequential progress is ideal (unlike
  Rarest-First).
• Lowest startup delay among policies evaluated.
• We believe this policy is optimal, but do not
  have a formal proof at this time. (Help!)

26
Model Validation Scenario

• Fluid simulation in ns-2
• Number of pieces M 100
• Range 100 to 200
• Piece size 128 KB
• Media playback rate r 2000 Kbps
• Download capacity 3200 Kbps (D 16 conns)
• Upload capacity 800 Kbps (U 4 conns)
• Range 300 Kbps to 2000 Kbps
• Arrival rate 50 (per media duration)
• Range 6.25 to 100.0
• Seed residence time 20 seconds
• Range 10 seconds to 150 seconds

27
Model Validation (1 of 3)

• Swarm population increases linearly with peer
  arrival rate
• In-Order(random) has higher system pop.
• In-Order(random) is sensitive to seed residence
  time, while other policies are not.

28
Model Validation (2 of 3)

• Download time improves when the seed residence
  time is increased
• Download time improves when the upload bandwidth
  is increased
• Good agreement with analytical models

29
Model Validation (3 of 3)

• In-Order(FCFS) achieves lowest startup delay
• Good agreement between simulation results and
  analytical model

30
Conclusions

• BitTorrent-like protocols can support scalable
  and efficient on-demand streaming.
• Rarest-First attains poor sequential progress.
• In-Order(FCFS) achieves best performance for
  media playback while attaining the same download
  latency of Rarest-First.
• Our analytic models accurately predict system
  performance (simulation validation).

31
Key Contributions

• Decoupling of download progress and sequential
  progress (key insight).
• Download latency and startup delay
  characterization for different policies.
• Effect of upload U and download D connections,
  not just total bandwidth.
• Distribution of download time and variability of
  progress (stream quality).
• Optimal structure for on-demand media streaming
  in P2P networks (?)

32
BitTorrent Model (optimal?)
Seed
Downloader
Seed
Downloader
U of upload connections C Per connection
throughput
Downloader
Torrent (with x downloaders and y seeds)
Arrival rate ?
Departure rate ? y
Peers (sorted by age)