Presentation is loading. Please wait.

Presentation is loading. Please wait.

Yang Guo Thomson Princeton Lab

Similar presentations


Presentation on theme: "Yang Guo Thomson Princeton Lab"— Presentation transcript:

1 Yang Guo Thomson Princeton Lab
Mesh or Multiple-Tree: A Comparative Study of P2P Live Streaming Approaches Nazanin Magharei, Reza Rejaie University of Oregon Yang Guo Thomson Princeton Lab 1/2/2019 Nazanin Magharei Infocom 2007

2 Nazanin Magharei Infocom 2007
Motivation Live P2P streaming has become increasingly popular approach for streaming live content to many receivers, IPTV e.g. SOPCAST, PPStream and TVUPlayer. Existing approaches can be divided into two classes: Tree-shaped overlay + push content delivery Mesh-shaped overlay + pull content delivery (swarming) Previous studies on live P2P streaming have often focused on design or evaluation of a particular mechanism. What are the differences and similarities between these two classes? How do the differences affect their performance? But this question remains unanswered 1/2/2019 Nazanin Magharei Infocom 2007

3 Overview: Live P2P Streaming
A live P2P streaming mechanism has two components: Overlay construction: How overlay is maintained, mesh or tree Content delivery: How overlay is used for content delivery, push or pull Goal: Maximizing delivered quality to peers in a scalable fashion Challenges: in-time delivery of packet despite churn, BW heterogeneity & asymmetry To cope with BW heterogeneity, stream is often encoded into multiple independent sub-streams, e.g. MDC 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 3 3

4 Multiple-tree approach
Overview Multiple-tree approach Basic idea: Maintain multiple diverse trees Each description of MDC content is pushed through a particular tree Each node joins a proper number of trees based on its incoming BW A common approach to maintain diverse stable trees (e.g. CoopNet, Splitstream) is to: Place a node as an internal node in only one tree and leaf node in all other trees Also try to maintain balanced and short trees Key design issue: tree construction SRC Tree 2 A Tree 1 D B C E F Key issue is tree construction and maintenance Mention that internal nodes can be distinguished by their colors from leaf nodes Internal node is placed as a child for the node with the lowest depth that has an empty slot Or has a child that is a leaf A node upon joining is placed at the first empty slot or if it is internal it can preempt an existing leaf node Might need to omit animation for leaf node joining D E F G A B C G 2 descriptions delivered to 7-peer overlay 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 4 4

5 Mesh-based approach Randomly connected uni-directional overlay mesh
Overview Mesh-based approach Randomly connected uni-directional overlay mesh Packet scheduling mech at each peer independently determines what packets to pull from individual parents Key design issue: packet scheduling at each peer SRC 2 1 3 5 4 6 7 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 5 5

6 This paper Illustrates the similarities & differences between tree- and mesh-based approaches to live P2P streaming, i.e. push- vs pull-based approaches. Examines to what extent their differences affect the performance of these two approaches, and why. Leveraging the notion of delivery tree per packet Our focus is on live P2P streaming 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 6 6

7 Similarities Overall shape of the resulting overlay is very similar
SRC Overlay construction: Overall shape of the resulting overlay is very similar Aggregate view of multiple trees is a directed mesh Content delivery: Peers receive different pieces of content through different paths Peer level: each peer receives content from multiple parents and sends it to multiple children System level: the collection of all edges used for delivery of a single packet to all peers form a source-rooted tree, called delivery tree 1 Tree 1 2 Tree 2 3 4 5 6 2 5 6 7 1 3 A 4 7 SRC 2 What does it mean ? In peer level .. 1 3 5 4 6 7 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 7 7

8 Differences Formation of delivery tree for individual packets
2 1 3 6 7 5 4 SRC Formation of delivery tree for individual packets Tree-based: Static per description Pinned down by overlay construction mechanism A drop in BW of a connection affects all the downstream peers Mesh-based: Dynamic per packet Direct result of packet scheduling at all peers How does a drop in BW of a connection affect other peers in a mesh? How is a delivery tree dynamically formed in a mesh-based approach? SRC 2 Don’t forget to mention that the dynamic formation is the effect of packet scheduling 1 3 5 4 6 7 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 8

9 PRIME: Delivery tree formation
Organized view: group peers into levels based on their shortest distance from source Delivery of packets has two phases: Diffusion phase Peers pull packets from parents in the higher level or source, these peers form a diffusion sub-tree Packets flow away from source Swarming phase Peers pull packets from parents in the same or lower levels SRC 1 2 Level 1 Diffusion direction 3 4 5 6 Level 2 Define diffusion subtree Now we saw that delivery tree consists of a diffusion subtree and a combination of swarming connection hanging from that 7 8 9 10 11 12 13 14 Level 3 Swarming direction 1/2/2019 Nazanin Magharei Infocom 2007

10 PRIME: Delivery tree formation
SRC 1 1 2 2 2 5 12 13 11 6 14 10 9 7 1 8 3 4 3 3 4 4 5 5 6 6 7 7 8 8 9 9 10 10 11 11 12 12 13 13 14 14 Lets take a closer look at an example of a delivery tree Mention many possible combination, see the paper for different types of swarming connections that can be attached to different locations Lets get back to that basic question of drop in BW of a connection .. Delivery tree has two components: whole or part of a diffusion subtree attached swarming connections Swarming connections can attach to the diffusion subtree in certain ways (see paper for details) 1/2/2019 Nazanin Magharei Infocom 2007

11 Evaluation Methodology
Many parameters could affect performance of P2P streaming mechanism overlay properties, peer dynamics, peer buffer. We adopt the following methodology to separate the effect of these parameters: Content delivery in static overlay (ns-2) Per-connection BW, Peer degree/number of trees, BW heterogeneity and Group size Content delivery in a distorted overlay (ns-2) Cohesion of the overlay structure under persistent churn (psim) Please see the paper for all the results 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 11 11

12 Extended effect of BW bottleneck
Results – Static overlay Extended effect of BW bottleneck % of connections (CDF) % of connections (CDF) Avg. BW utilization (%) , Tree-based, degree=8 Avg. BW utilization (%) , Mesh-based, degree=8 In tree, lower level connections have a lower average utilization  a BW bottleneck in upper levels has an extended effect on all downstream connections In mesh, connections regardless of their location have a high BW utilization (>95%) 1/2/2019 Nazanin Magharei Infocom 2007

13 Peer degree Results – Static overlay
Avg. BW utilization (%) Degree Increasing degree in tree reduces depth  reduces the probability of cumulative effect upstream bottleneck Increasing degree in mesh-based approach increases diversity among parents 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 13 13

14 Pattern of content delivery
Results – Static overlay Pattern of content delivery % of population (CDF) % of population (CDF) Avg. hop count – Mesh-based Avg. hop count – Tree-based Avg. hop count in mesh is larger due to the flexibility of packet scheduling to receive a packet from longer paths For both approaches, increasing peer degree decreases avg. hop count but for different reasons: In mesh-based approach due to the increase in diversity In tree-based approach due to the decrease in depth 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 14 14

15 Nazanin Magharei Infocom 2007
Results – Dynamic overlay Distorted overlay 5 2 12 10 Avg. BW utilization (%) 4 Source 7 1 3 8 11 6 9 13 % of departed peers Distorted overlay represents worst case connectivity with churn In tree, due to the diverse placement of peers in trees, departure of any peer reduces the quality of all of its descendants In mesh, flexibility in utilizing swarming connections minimize the impact of peer departure 1/2/2019 Nazanin Magharei Infocom 2007

16 Stability of path from source
Results – Dynamic overlay Stability of path from source Avg. interval between ancestor change (sec) Avg. interval between ancestor change (sec) Group size – Mesh-based Group size – Tree-based Stability of path from source Increasing group size, decreases stability of path from source in both approaches due to an increase in avg. distance of peers. Long-lived peers observe a higher degree of stability among their ancestor in mesh Long-lived peers gradually move to higher levels and increase stability Paths from source are more stable in mesh 1/2/2019 Nazanin Magharei Infocom 2007

17 Conclusions Mesh-based approach consistently exhibits a superior performance over tree-based Mesh is able to effectively utilize available resources as the overlay grows mainly due to the dynamic mapping of content to parents – tree can not Peer departure has a local effect on a mesh but it always affect a sub-tree (where it is internal node) in a tree because of the diverse placement of peers Ongoing work: examining other tree construction algorithms, especially those that dynamically change parents to improve performance, e.g. ChunkySpread. For further information on our projects on P2P streaming visit 1/2/2019 1/2/2019 Nazanin Magharei Infocom 2007 Nazanin Magharei Infocom 2007 17 17


Download ppt "Yang Guo Thomson Princeton Lab"

Similar presentations


Ads by Google