1. Understanding Mesh-based Peer-to-Peer Streaming Nazanin Magharei Reza Rejaie

2. Motivation & Related works P2P overlays support one-to-many multimedia streaming applications without any special network support P2P streaming mechanism goals are:  Maximizing delivered quality  Achieving scalability  Accommodating heterogeneity and asymmetry of peers’ bandwidth  Providing resiliency despite peers’ dynamic Common approach is tree-based overlay construction with push-based content delivery  Delivered quality is limited to each peer  Leaf peers’ bandwidth can not be utilized

3. Mesh-based streaming Mesh-based overlay construction with pull- based content delivery  Inspired by file-swarming mechanisms (e.g. BitTorrent( Distribute pieces of a file among different peers Most peers contribute their outgoing bandwidth Scalable Resilient to peers’ dynamic

4. Challenges & Goals Problem: Incorporating swarm-like content delivery mechanism to support live media streaming to a large number of dynamic peers Challenges:  In-time requirement of content delivery  Limited availability of future content Goals:  Effectively utilizing outgoing bandwidth of all participating peers  Maximizing delivered quality to each peers  Minimizing buffer (playback delay from source) In this paper we try to answer these fundamental questions:  How do the key properties of the overlay (e.g. degree) affect the delivered quality to individual peers?  What is the global flow pattern of content over mesh-based overlay to achieve system goals?

5. Mesh-based P2P Streaming Overview Two key components:  Overlay Construction Randomly connected and directed mesh Each peer has multiple parents and multiple children  Simple  Resilient to churn  Content Delivery Push reporting by parents Periodically pull requesting by child peers New segment of length  generated by source every  sec. Peers delay their playout time by at least  behind source MDC stream to deliver maximum quality to each peer based on its available bandwidth 2 1 3 6 7 5 4 12 13 9 10 11 8 Source

6. Performance Bottleneck To separate causes of low quality, define two performance bottlenecks: Bandwidth Bottleneck  Avg. BW for a connection between parent p and child c: MIN (outbw p /outdeg p, inbw c /indeg c )  If these two ratios are not equal, one of them causes BW bottleneck.  Condition : outbw p /outdeg p = inbw c /indeg c = bwpf  All connections in the overlay must have roughly the same bandwidth. Content Bottleneck  Useful content among some parents of a given peer is not sufficient to fully utilize its available bandwidth c p outdeg p indeg c

7. 213 6754 [t 0,t 0 +  ][t 0  t 0 +2  ][t 0  t 0 +3  ] Content Bottleneck cont, Data units = bwpf *  (interval) Each parent peer should have at least one useful data unit per Interval  for each of its child peer Unique data units of each segment must be delivered to peers within  intervals after its generation time by source Buffer    delay  Goal is to:  minimize content bottleneck  maximize quality  minimize    minimize playback delay To achieve these goals content delivery should have two phases:  Diffusion  swarm [t0]

8. Organized view of the overlay mesh Peers organized into levels, based on their shortest distance from source Each peer with degree deg in level n has one parent in level n-1 (diffusion parent) and deg-1 parents in the same or lower level (swarming parents)  Peer 13 in level 3 has one parent 7 in level 2 the rest in level 3 Population of levels increases by going down into levels : pop(n) <= deg src * deg (n-1) Level 1 Level 2 Level 3 2 13 6 75 4 12 13 9 10 11 8 Source

9. Content Delivery: Diffusion Diffusion phase  Fastest time for pulling all data units of a segment from source to the lowest level = depth  sec  All connections from diffusion parents should be exclusively used for diffusion of new data units Each subtree rooted at level 1 peers receives the same data unit which we called diffusion subtree After depth intervals each peer has one data units of the segment 2 13 6 75 4 12 13 9 10 11 8 Source

10. Content Delivery: Swarm Swarming phase  Pulling the missing data units of the segment from swarming parents  Only a swarming parent in a different diffusion subtree can rapidly provide a new useful data units Swarming parent 11 for child 12 are at the same diffusion subtree ! Even larger for parent 8 to child 1, 3 swarming intervals Swarming phase for a segment may take more than 1 interval depends on the location of swarming parents. K min = minimum number of swarming intervals for which 90% of peers quality > 90% K min =  min – depth 2 13 6 75 4 12 13 9 10 11 8 Source

11. Key overlay property Goal : maximizing quality while minimizing delay Increasing degree can help to achieve goals by :  reducing depth and  increasing diversity of swarm connections But increases observed loss rate thus decreases throughput of each connection  Limited degree range for a bandwidth? What is a good degree for a peer with certain bandwidth? That is what is a good bwpf ?

12. Evaluation K min is fixed =3 Very low degree, decrease in diversity of swarm parents  increase in swarming interval Very high degree, increase in loss rate  increase in swarming interval Higher bw scenario, higher bwpf ratio There is a sweet range of bwpf values

13. Results Depth gradually increases Swarming interval K min does not change with peer population as number of diffusion Subtrees are fixed so the diversity is fixed  scales well Increasing degree, increases transmission rate

14. Conclusions future works We presented a global pattern of content delivery that is able to effectively utilize outgoing bw of peers while maximizing delivered quality with minimum buffer requirement There is a sweet range of peer degree (bwpf) that maximizes delivered quality to individual peers with minimum buffer. Future work: Conducting a comprehensive evaluation of P2P mesh- based streaming protocol.  Heterogeneity in peer properties (e.g. access link bandwidth)  Peer behavior (e.g. packet scheduling strategy)  Churn