1. University of Würzburg Distributed Systems Prof. Dr. P. Tran-Gia An Application-level Active Networks Based Architecture for the Performance Management of Peer-to-Peer Services H. deMeer and K. Tutschku OPENSIG 2001 Workshop: „Next Generation Network Programming“ London, 24./25. September 2001

2. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Overview  What is Peer-to-Peer Networking?  Selected P2P Architectures  Resource and Performance Management of P2P Services  Goals and Approaches  ALAN-based Architecture for P2P Performance Management  ALAN superpeers  superpeer implementation  Conclusion

3. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Peer-to-Peer: Introduction  Tremendous industrial hype and strong research attention:  Napster: 40 million user deployments in two years  O’Reilly P2P conference, March 2001 / panel at InfoCom 2001  Typical applications: –file sharing (Napster, Gnutella) –group collaboration (Jxta [Sun], Groove) –distributed storage (PAST/Farsite [Microsoft], Chord [MIT]) –distributed computation (SETI@home)

4. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Peer-to-Peer: Features  Basic features:  distributed architecture –composed of voluntary and ad-hoc membership of peers –symmetric roles (serving, downloading, routing) throughout system: servent = server + client –weak connectivity: handles variable connectivity as the norm available at the edges of the Internet  Hope:  instant services, no cost of administration, re-use of resource  quality and robustness scales with size of infrastructure  Many many challenges:  trust, semantics, location, efficiency/performance, operation, robustness

5. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Selected P2P Applications - SETI@home  Purpose:  uses idle CPU cycles on ordinary PCs for massively parallel analysis of extraterrestrial radio signals  Architecture:  central SETI@home server distributes data  analyses done locally by SETI@home screen saver  Classification (according K. Kant):  (scattered, organized, isolated, non_RT)  Similar architectures:  Napster

6. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Selected P2P Applications - Gnutella (1)  Purpose:  distributed and anonymous file sharing  Servents operate completely without central control  Exploits unused storage on edge nodes  Characteristics:  message broadcasting for node discovering and search requests;  forming of overlay network; connecting: join the “several known hosts”  user data transfer: store and forward using HTTP  Classification:  (scattered, scattered, isolated, non_RT)

7. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Selected P2P Applications - Gnutella (2)  flooding of PING/PONG messages; broadcasting range limited by TTL counter.  short time memory of messages already seen; prevents re- broadcasting; GUIDs to distinguish msg G-Node Ping Pong Ping  Call-and-Response protocol mechanism: node discovery G-Node

8. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Selected P2P Applications - Gnutella (3) 1) Node A asks Node B for data. 3) B forwards the request to its neighbors. 4) These return any match- ing info. 5) B looks up source of request. 6) B returns matching info G-Node A G-Node A G-Node B G-Node B G-Node C G-Node C G-Node D G-Node D  Call-and-Response protocol mechanism: search query / download 2) B keeps a record that A initiated the request Node A 7) A may initiate download using HTTP  search: Query/Query-Response (flooding!)  download: GET/PUSH. (direct transmission)

9. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Selected P2P Applications - Gnutella (4)  Limitations/Difficulties:  unstable/loose connectivity of the servents  performance management difficult  scalability: e.g. TTL=10, every node broadcasts to six others  msg; problem in huge networks  low TTL, low search horizon  denial-of-service attacks  Challenges:  Robustness –availability of resources; hard to predict the consequences of failures; administrative actions  Performance –bandwidth consumption; scalability; end-to-end quality-of-service; overload / network planning

10. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku P2P Resource/Performance Management  Goals:  maximizes a peer’s utility to the overall system while minimizing its potential threat  increase stability  introduce administrative rules  Problems to tackle:  resource discovery/allocation  reduction of synchronization traffic  aggregation and self-organization  simple overuse (e.g. freeloaders)  bandwidth/latency/packet loss

11. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku P2P Resource/Performance Management  Approaches:  enhance P2P protocols: –topology construction (JXTA) –sophisticated group multicasting (Lee et al., 2001)  accountability: –used in Free Haven project / Mojo Nation –restricting access: “micro payments” –selecting favored users: reputation system  superpeers: –Morpheus/KaZaA, Clip2’s Gnutella Reflector –Virtual Active Peer based on Application-Level Active Networks

12. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku “Virtual Active Peer” Architecture  Superpeer:  layering provides control capability  facilitates limited topology management  Application-level Routing  optimizes for different metrics (e.g. privacy, policies, latency)  provides smart multicast, caching and replication capabilities

13. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Application Level Active Networks (ALAN)  proposed by Fry and Ghosh [1999]  active elements on application level  system consists of proxylets and EEPs:  proxylets: –dynamic code modules specifying the protocol handling –single copy stored on central server; identified by reference (URL)  EEPs: –execution environment for proxylets; located at strategic points –proxylets ctrl: load, run, modify, stop –dynamic deployment –system proxylets: routing, discovery, and error handling

14. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku ALAN – Architecture Execution Environment for Proxylets (EEP) Monitor Interface Protocol Server Gnutella Active Peer Proxylet Gnutella Ctrl Interface Control Interface Proxylet Server Gnutella Mon. Interface Peer Super- peer Proxylet

15. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Superpeer Architecture  Application Optimization Layer:  management of peer-to-peer relation on application level; using application level performance metrics  enforcement of connections with predictable performance  Virtual Control Cache:  application-level aggregation capability  facilitates differentiation  Network Optimization Layer:  optimal mapping wrt. transport network capabilities  enables dynamic traffic engineering Gnutella Active Proxylet Application Optimization Layer Virtual Ctrl Cache Network Optimization Layer

16. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Superpeer Architecture  Topology control:  set-up of stable connec- tions to other EEPs  instantiation of new active peers at other EEPs  termination of Gnutella connections  distribution of announce- ments (ping msg) Gnutella Active Proxylet Application Optimization Layer Virtual Ctrl Cache Network Optimization Layer Topology CtrlPolicy Ctrl Performance Monitoring  Policy control:  user, network, or servent based  TTL / connectivity  prioritization / localiza- tion  Perf. Mon:  query traffic  ping/pong traffic  robustness of connectivity  response time  throughput

17. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku P2P JTella Dependencies GNUTellaConnection host, port Router ConnectionList, HostCache IncomingConnectionManager Router, ConnectionList, ConnectionData OutgoingConnectionManager Router, ConnectionList, ConnectionData KeepAliveThread ConnectionData HostCache HostFeed SearchMonitorSession MessageReceiver OriginateTable BoundedQueue messageQueue RouteTable queryhit RouteTable query RouteTable pings Vector searchReceivers Vector pushReceivers MessageReceiver NodeConnection RouteMessage Message esp. Ping-, Pong-, Push-, Query- and QueryReplyMessage ServerSocket HostCache StarterPool Gnutella Proxylet (Application Optimization Layer)  Active Peer Proxylet:  based on JTella 0.6 API  full Gnutella routing capability  layered architecture  on-going implementation

18. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Self-Organization  active peers topology control permits local self-organization  e.g. accepting only peers with similar response time  forming of zones of similar performance  proxylet ctrl permits initiation of new active peers  superpeer architecture able adapt to to spatially varying performance condition

19. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Approach to the Management of Wireless Ad-hoc Networks  Wireless ad-hoc networks:  highly distributed architecture –weak relationship between participating nodes –highly varying link performance  constantly changing set of moving/dis- appearing nodes  highly dependent on routing  forming of zones of similar performance  based on radio link performance metrics

20. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku Conclusion & Outlook  P2P networking is a promising paradigm for services operating at the edges of the network  weak connectivity of P2P services challenges the resource and performance management  ALAN-based “Virtual Active Peer” approach  combines the advantages of distributed and centralized ctrl architectures  facilitates methods to increase service stability and service performance  permits the forming of zones of equal performance  Outlook:  performance management schemes require new models –networks dynamics: what is best node to connect to? residual time of servents in the system? –combined synchronization/user data traffic model?

21. University of Würzburg Distributed Systems Performance Management Architecture for P2P Services H. DeMeer / K. Tutschku End of talk Thank you Q & A