multi-torrent: a performance study - Semantic Scholar
MULTI-TORRENT: A PERFORMANCE STUDY Yan Yang, Alix L.H. Chow, Leana Golubchik Internet Multimedia Lab University of Southern California
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
2 Internet Multimedia Lab
BACKGROUND AND INTRODUTION(1) BitTorrent
(BT) overview
An extremely popular P2P file sharing application. Nodes request chunks of data from their neighbors after it receives “start-up” information from Tracker. Nodes are termed “leechers” or “seeds” based on if they have complete copy of the data file. Each leecher picks a number of nodes to upload file chunks. (unchoke) A subset of the nodes (typically 4) are unchoke using Titfor-Tat (TFT). (regular unchoke) A subset of the nodes (typically 1) is picked randomly. (optimistic unchoke)
Each seed also picks a subset of neighbors (typically 5) to upload.
Internet Multimedia Lab
3
BACKGROUND AND INTRODUCTION(2) Why
we study multi-torrent?
Most previous works focus on a single torrent. [4, 5, 6 ] In practice, many published torrents are related. Different episodes of a TV show, movies/music form the same genre, etc.
Previous study shows such related interests result in most peers (> 85%) participating in multi-torrents. [1] We believe there is a reasonably high probability that users would download multi-torrents simultaneously. Users get start-up information from a particular source. There is correlation in contents.
However, current BT system doesn’t relates these torrent together.
Internet Multimedia Lab
4
BACKGROUND AND INTRODUCTION(3) Advantages
Help newly joined nodes ramp up faster. Help nodes nearing end of their downloads find the last few file chunks faster. (end-game behavior) Keep a torrent “alive”.
Our
of multi-torrent
contributions
Propose a multi-torrent BT system which can be easily implemented through fairly small modification to the current BT protocol. A “cross-torrent-based” TFT is proposed to provide incentives to nodes as seeds. Extensive simulate-based study shows: Multi-torrent improves the overall performance. Multi-torrent provides incentives for nodes to act as seeds.
Internet Multimedia Lab
5
PROPOSED APPROACH(1) A
motivational example
6 Internet Multimedia Lab
PROPOSED APPROACH(2)
Cross-Torrent TFT (CTFT)
Peers’ downloading rate in all participating torrent are aggregated. Aggregation is done in a weighted manner. Higher weight (> 1) given to the downloading rate from seeding torrents. Normal weight (=1) given to the downloading rate from leecher torrents.
Total contribution of peer Nx to a peer Ny is used to rank peers for TFT unchoking by node Nx.
Download rate of node Nx from node Ny
#Torrents
∑ w (y) × D (x, y) i
i
i=1
w i (y) > 1 N y is seed on torrent i otherwise w i (y) = 1 Internet Multimedia Lab
€
7
PROPOSED APPROACH(3) CTFT-DS
Motivated by unbalanced #seeds when there is sufficient heterogeneity in the file sizes of different torrents. CTFT-DS is used to mitigate this problem.
Each nodes does a local estimate of the ratio of seeds to leechers. Only participate as seed in those torrents where seed to leecher ratio is below R. This local estimate and which torrents to seed is reevaluated every S time units. R and S are system parameters.
8 Internet Multimedia Lab
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
9 Internet Multimedia Lab
SIMULATION SETUP(1) Experiment
setup
Node
class
Parameters
Value
File Size
200MB
Simulation Time
12 hours (+3 hours warm-up)
Average node inter-arrival time
45 sec
Peer set size
40
Leccher unchoking
4 Regular + 1 optimistic
#Seed unchoking
5
Class
Fraction
Download Cap
Upload Cap
Slow
40%
1500 kbps
128 kbps
Fast
60%
5000 kbps
512 kbps
10 Internet Multimedia Lab
SIMULATION SETUP(2)
Simulator An event based simulator in [2] with extension. The system starts with 1 original per torrent and it stays in the system for the duration of simulation. All experiments use same node arrival sequence. All experiments use same torrent selection sequence in case of random torrent selection. Seed uploading algorithm is updated to be like latest BT protocol.
Performance metrics Average download time (over all torrents) Average download time (last torrents)
11 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(1)
Different number of torrents
Fixed selection of torrents Random selection of torrents
12 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(2)
Node Inter-Arrival Time
Homogeneous system with only fast nodes. Each node joins same 3 torrents.
13 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(3)
Different fraction of node staying around
Each node randomly select 3 torrents to join out of 10. Faction of staying nodes varies from 20% to 80%.
14 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(4)
Different CTFT Weight
Randomly select 3 torrents to join out of 10. Each node stays around with probability 0.5.
15 Internet Multimedia Lab
RESULTS: APPLICATION
An online game patch system
File sizes are 492MB, 256MB, and 80MB.
based on WoW major patch release 2.0, 2.1, and 2.2 [3].
Homogeneous system with fast nodes only. Each node participates all 3 torrents.
16 Internet Multimedia Lab
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
17 Internet Multimedia Lab
DISCUSSION Performance
We expect even bigger improvement in real world.
System
in the wild
parameters
Exploration of various system and proposed schemes’ parameters is our on-going effort.
Malicious
behavior
A malicious node can take advantage of the weight of CTFT, how to detect malicious node is one of our future works. How malicious behavior affects the system performance is one of our future works.
18 Internet Multimedia Lab
CONCLUSION Performance
gains are possible through multi-
torrent. Multi-torrent
is a promising research area with remaining future directions.
Thank You!
Internet Multimedia Lab
19
REFERENCE 1.
2.
3. 4.
5.
6.
Guo, Chen, Xiao, Tan, Ding, and Zhang, Measurements, analysis and modeling of bittorrentlike systems. In IMC 2005. Bharambe, Herley, and Padmanabhan, Analyzing and improving bittorrent performance. In INFOCOM 2006. Blizzard. World of warcraft website. Legout, Liogkas, Kohler, and Zhang, Clustering and sharing incentives in bittorrent system. In SIGMETRICS 2007. Piatek, Isdal, Anderson, Krishnamurthy, and Venkataramani, Do Incentives build robustness in bittorrent? In NSDI 2006. Fan, Chiu, and Lui, The delicate tradeoffs in bittorrent-like file sharing protocol design. In ICNP 2006.
Internet Multimedia Lab
20
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
2 Internet Multimedia Lab
BACKGROUND AND INTRODUTION(1) BitTorrent
(BT) overview
An extremely popular P2P file sharing application. Nodes request chunks of data from their neighbors after it receives “start-up” information from Tracker. Nodes are termed “leechers” or “seeds” based on if they have complete copy of the data file. Each leecher picks a number of nodes to upload file chunks. (unchoke) A subset of the nodes (typically 4) are unchoke using Titfor-Tat (TFT). (regular unchoke) A subset of the nodes (typically 1) is picked randomly. (optimistic unchoke)
Each seed also picks a subset of neighbors (typically 5) to upload.
Internet Multimedia Lab
3
BACKGROUND AND INTRODUCTION(2) Why
we study multi-torrent?
Most previous works focus on a single torrent. [4, 5, 6 ] In practice, many published torrents are related. Different episodes of a TV show, movies/music form the same genre, etc.
Previous study shows such related interests result in most peers (> 85%) participating in multi-torrents. [1] We believe there is a reasonably high probability that users would download multi-torrents simultaneously. Users get start-up information from a particular source. There is correlation in contents.
However, current BT system doesn’t relates these torrent together.
Internet Multimedia Lab
4
BACKGROUND AND INTRODUCTION(3) Advantages
Help newly joined nodes ramp up faster. Help nodes nearing end of their downloads find the last few file chunks faster. (end-game behavior) Keep a torrent “alive”.
Our
of multi-torrent
contributions
Propose a multi-torrent BT system which can be easily implemented through fairly small modification to the current BT protocol. A “cross-torrent-based” TFT is proposed to provide incentives to nodes as seeds. Extensive simulate-based study shows: Multi-torrent improves the overall performance. Multi-torrent provides incentives for nodes to act as seeds.
Internet Multimedia Lab
5
PROPOSED APPROACH(1) A
motivational example
6 Internet Multimedia Lab
PROPOSED APPROACH(2)
Cross-Torrent TFT (CTFT)
Peers’ downloading rate in all participating torrent are aggregated. Aggregation is done in a weighted manner. Higher weight (> 1) given to the downloading rate from seeding torrents. Normal weight (=1) given to the downloading rate from leecher torrents.
Total contribution of peer Nx to a peer Ny is used to rank peers for TFT unchoking by node Nx.
Download rate of node Nx from node Ny
#Torrents
∑ w (y) × D (x, y) i
i
i=1
w i (y) > 1 N y is seed on torrent i otherwise w i (y) = 1 Internet Multimedia Lab
€
7
PROPOSED APPROACH(3) CTFT-DS
Motivated by unbalanced #seeds when there is sufficient heterogeneity in the file sizes of different torrents. CTFT-DS is used to mitigate this problem.
Each nodes does a local estimate of the ratio of seeds to leechers. Only participate as seed in those torrents where seed to leecher ratio is below R. This local estimate and which torrents to seed is reevaluated every S time units. R and S are system parameters.
8 Internet Multimedia Lab
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
9 Internet Multimedia Lab
SIMULATION SETUP(1) Experiment
setup
Node
class
Parameters
Value
File Size
200MB
Simulation Time
12 hours (+3 hours warm-up)
Average node inter-arrival time
45 sec
Peer set size
40
Leccher unchoking
4 Regular + 1 optimistic
#Seed unchoking
5
Class
Fraction
Download Cap
Upload Cap
Slow
40%
1500 kbps
128 kbps
Fast
60%
5000 kbps
512 kbps
10 Internet Multimedia Lab
SIMULATION SETUP(2)
Simulator An event based simulator in [2] with extension. The system starts with 1 original per torrent and it stays in the system for the duration of simulation. All experiments use same node arrival sequence. All experiments use same torrent selection sequence in case of random torrent selection. Seed uploading algorithm is updated to be like latest BT protocol.
Performance metrics Average download time (over all torrents) Average download time (last torrents)
11 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(1)
Different number of torrents
Fixed selection of torrents Random selection of torrents
12 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(2)
Node Inter-Arrival Time
Homogeneous system with only fast nodes. Each node joins same 3 torrents.
13 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(3)
Different fraction of node staying around
Each node randomly select 3 torrents to join out of 10. Faction of staying nodes varies from 20% to 80%.
14 Internet Multimedia Lab
RESULTS: DESIGN SPACE EXPLORATION(4)
Different CTFT Weight
Randomly select 3 torrents to join out of 10. Each node stays around with probability 0.5.
15 Internet Multimedia Lab
RESULTS: APPLICATION
An online game patch system
File sizes are 492MB, 256MB, and 80MB.
based on WoW major patch release 2.0, 2.1, and 2.2 [3].
Homogeneous system with fast nodes only. Each node participates all 3 torrents.
16 Internet Multimedia Lab
OVERVIEW Background
and Introduction Proposed Approach Simulation Study Discussion Conclusion
17 Internet Multimedia Lab
DISCUSSION Performance
We expect even bigger improvement in real world.
System
in the wild
parameters
Exploration of various system and proposed schemes’ parameters is our on-going effort.
Malicious
behavior
A malicious node can take advantage of the weight of CTFT, how to detect malicious node is one of our future works. How malicious behavior affects the system performance is one of our future works.
18 Internet Multimedia Lab
CONCLUSION Performance
gains are possible through multi-
torrent. Multi-torrent
is a promising research area with remaining future directions.
Thank You!
Internet Multimedia Lab
19
REFERENCE 1.
2.
3. 4.
5.
6.
Guo, Chen, Xiao, Tan, Ding, and Zhang, Measurements, analysis and modeling of bittorrentlike systems. In IMC 2005. Bharambe, Herley, and Padmanabhan, Analyzing and improving bittorrent performance. In INFOCOM 2006. Blizzard. World of warcraft website. Legout, Liogkas, Kohler, and Zhang, Clustering and sharing incentives in bittorrent system. In SIGMETRICS 2007. Piatek, Isdal, Anderson, Krishnamurthy, and Venkataramani, Do Incentives build robustness in bittorrent? In NSDI 2006. Fan, Chiu, and Lui, The delicate tradeoffs in bittorrent-like file sharing protocol design. In ICNP 2006.
Internet Multimedia Lab
20
