1 A Hybrid Architecture for Cost-Effective On-Demand Media Streaming Mohamed Hefeeda & Bharat Bhargava CS Dept, Purdue University Support: NSF, CERIAS FTDCS 2003

2  Imagine a media server streaming movies to clients -Target environments Distance learning, corporate streaming, …  Current approaches -Unicast Centralized Proxy caches CDN (third-party) -Multicast Network layer Application layer  Proposed approach (Peer-to-Peer) Motivations

3 Unicast Streaming: Centralized  Pros -Easy to deploy, administer  Cons -Limited scalability -Reliability concerns -Load on the backbone -High deployment cost $$$…..$  Note: -A server with T3 link (~45 Mb/s) supports only up to 45 concurrent users at 1Mb/s CBR!

4 Unicast streaming: Proxy  Pros -Better performance -Less load on backbone and server -Prefix caching, selective caching, staging, …  Cons -Proxy placement and management -High deployment cost $$$…..$

5 Unicast Streaming: CDN  Pros -Good performance -Suitable for web pages with moderate-size objects  Cons -Co$t: CDN charges for every megabyte served!  -Not suitable for VoD service; movies are quite large (~Gbytes)  Note [Raczkowski’02] -Cost ranges from 0.25 to 2 cents/MByte -For a one-hour streamed to 1,000 clients, content provider pays $264 to CDN (at 0.5 cents/MByte)!

6 Multicast Streaming: Network Layer  Pros -Efficient! -Asynchronous client Patching, skyscraper, … [e.g., Mahanti, et al. ToN’03]  Cons -Scalability of the routers -Asynchronous  tune to multiple channels -Require high inbound bandwidth (2R) -Not widely deployed Patch stream

7 Multicast Streaming: Application Layer  Pros -Deployable  Cons -Assumes end systems can support (outbound bandwidth) multiple folds of the streaming rate

8 P2P Approach: Key Ideas  Make use of underutilized peers’ resources  Make use of heterogeneity  Multiple peers serve a requesting peer  Network-aware peers organization

9 P2P Approach (cont’d)  Pros -Cost effective -Deployable (no new hardware, nor network support) -No extra inbound bandwidth (R) -Small load on supplying peers (< R) -Less stress on backbone  Challenges -Achieve good quality Unreliable, limited capacity, peers Highly dynamic environment

10 P2P Approach: Entities  Peer -Level of cooperation (rate, bw, coonections)  Super peer -Help in searching and dispersion  Seed peer -Initiate the streaming  Stream -Time-ordered sequence of packets  Media file -Divided into equal-size segments  Super Peers play special roles  Hybrid system

11 Hybrid System: Issues  Peers Organization -Two-level peers clustering -Join, leave, failure, overhead, super peer selection  Operation -Client protocol: manage multiple suppliers Switch suppliers Effect of switching on quality  Cluster-Based Searching -Find nearby suppliers  Cluster-Based Dispersion -Disseminate new media files Details are in the extended version of the paper

12 Peers Organization: Two-Level Clustering  Based on BGP tables -[Oregon RouteViews]  Network-cluster [KW 00] -Peers sharing the same network prefix /16 (purdue), /16 (cmu) , , ,  AS-cluster -All network clusters within an Autonomous System Snapshot of a BGP routing table

13 Peers Organization: Join Bootstrap data structure Join

14 Client Protocol  Building blocks of the protocol to be run by a requesting peer  Three phases -Availability check Search for all segments with the full rate -Streaming Stream (and playback) segment by segment -Caching Store enough copies of the media file in each cluster to serve all expected requests from that cluster Which segments to cache?

15 Client Protocol (cont'd)  Phase II: Streaming t j = t j-1 + δ /* δ: time to stream a segment */ For j = 1 to N do At time t j, get segment s j as follows: P jConnect to every peer P x in P j (in parallel) and Download from byte b x-1 to b x -1 Note: b x = |s j | R x /R Example: P 1, P 2, and P 3 serving different pieces of the same segment to P 4 with different rates

16 Dispersion (cont'd)  Dispersion Algorithm (basic idea): -/* Upon getting a request from P y to cache N y segments */ -C  getCluster (P y ) -Compute available (A) and required (D) capacities in cluster C -If A < D P y caches N y segments in a cluster-wide round robin fashion (CWRR) –All values are smoothed averages –Average available capacity in C: –CWRR Example: (10-segment file) P 1 caches 4 segments: 1,2,3,4 P 2 then caches 7 segments: 5,6,7,8,9,10,1

17 Evaluation Through Simulation  Client parameters -Effect of switching on quality  System Parameters -Overall system capacity, -Average waiting time, -Load/Role on the seeding peer -Scenarios: different client arrival patterns (constant rate, Poisson, flash crowd) and different levels of peer cooperation  Performance of the dispersion algorithm -Compare against random dispersion algorithm

18 Simulation: Topology –Large, Hierarchical (more than 13,000 nodes) -Ex. 20 transit domains, 200 stub domains, 2,100 routers, and a total of 11,052 end hosts –Used GT-ITM and ns-2

19 Client: Effect of Switching  Initial buffering is needed to hide suppliers switching  Tradeoff: small segment size  small buffering but more overhead (encoding, decoding, processing)

20 System: Capacity and Load on Seed Peer Load on seeding peer ─ The role of the seeding peer is diminishing For 50%: After 5 hours, we have 100 concurrent clients (6.7 times original capacity) and none of them is served by the seeding peer System capacity

21 System: Under Flash Crowd Arrivals  Flash crowd  sudden increase in client arrivals

22 System: Under Flash Crowd (cont'd) ─ The role of the seeding peer is still just seeding During the peak, we have 400 concurrent clients (26.7 times original capacity) and none of them is served by the seeding server (50% caching) Load on seeding peer System capacity

23 Dispersion Algorithm ─ Avg. number of hops: 8.05 hops (random), 6.82 hops (ours)  15.3% savings ─ For a domain with a 6-hop diameter: Random: 23% of the traffic was kept inside the domain Cluster-based: 44% of the traffic was kept inside the domain 5% caching

24 Conclusions  A hybrid model for on-demand media streaming -P2P streaming; Powerful peers do more work -peers with special roles (super peers) -Cost-effective, deployable -Supports large number of clients; including flash crowds  Two-level peers clustering -Network-conscious peers clustering -Helps in keeping the traffic local  Cluster-based dispersion pushes contents closer to clients (within the same domain)  -Reduces number of hops traversed by the stream and the load on the backbone

25 Thank You!

26 P2P Systems: Basic Definitions  Peers cooperate to achieve desired functions -Cooperate: share resources (CPU, storage, bandwidth), participate in the protocols (routing, replication, …) -Functions: file-sharing, distributed computing, communications, …  Examples -Gnutella, Napster, Freenet, OceanStore, CFS, CoopNet, SpreadIt, …  Well, aren’t they just distributed systems? -P2P == distributed systems? Background

27 P2P vs. Distributed Systems  P2P = distributed systems++; -Ad-hoc nature -Peers are not servers [Saroui et al., MMCN’02 ] Limited capacity and reliability -Much more dynamism -Scalability is a more serious issue (millions of nodes) -Peers are self-interested (selfish!) entities 70% of Gnutella users share nothing [Adar and Huberman ’00] -All kind of Security concerns Privacy, anonymity, malicious peers, … you name it! Background

28 P2P Systems: Rough Classification [Lv et al., ICS’02], [Yang et al., ICDCS’02]  Structured (or tightly controlled, DHT) +Files are rigidly assigned to specific nodes +Efficient search & guarantee of finding –Lack of partial name and keyword queries Ex.: Chord [Stoica et al., SIGCOMM’01], CAN [Ratnasamy et al., SIGCOMM’01], Pastry [Rowstron and Druschel, Middleware’01]  Unstructured (or loosely controlled) +Files can be anywhere +Support of partial name and keyword queries –Inefficient search (some heuristics exist) & no guarantee of finding Ex.: Gnutella  Hybrid (P2P + centralized), super peers notion) -Napster, KazaA Background

29 File-sharing vs. Streaming  File-sharing -Download the entire file first, then use it -Small files (few Mbytes)  short download time -A file is stored by one peer  one connection -No timing constraints  Streaming -Consume (playback) as you download -Large files (few Gbytes)  long download time -A file is stored by multiple peers  several connections -Timing is crucial Background

30 Related Work: Streaming Approaches  Distributed caches [e.g., Chen and Tobagi, ToN’01 ] -Deploy caches all over the place -Yes, increases the scalability Shifts the bottleneck from the server to caches! -But, it also multiplies cost -What to cache? And where to put caches?  Multicast -Mainly for live media broadcast -Application level [Narada, NICE, Scattercast, Zigzag, … ] -IP level [e.g., Dutta and Schulzrine, ICC’01] Widely deployed? Background

31 Related Work: Streaming Approaches  P2P approaches -SpreadIt [Deshpande et al., Stanford TR’01] Live media  Build application-level multicast distribution tree over peers -CoopNet [Padmanabhan et al., NOSSDAV’02 and IPTPS’02] Live media  Builds application-level multicast distribution tree over peers On-demand  Server redirects clients to other peers  Assumes a peer can (or is willing to) support the full rate  CoopNet does not address the issue of quickly disseminating the media file Background