Internet Real-Time Laboratory Wing Ho (Andy) Yuen Columbia University

1. IRT – Internet Real-Time Lab • Affiliated with the Computer Science Department at Columbia University • conducts research in the areas of Internet and multimedia services: Internet telephony, wireless and mobile networks, streaming, quality of service, resource reservation, dynamic pricing for the Internet, network measurement and reliability, service location, network security, media on demand, content distribution networks, multicast networks and ubiquitous and context-aware computing and communication. Project Title • Performance Evaluation of Time-based and Hop-based TTL Schemes in 7DS Systems • Qualitative Comparison of TB and HB • For >>, E[M]bmax, Hop-based scheme has storage cost and storage-time cost independent of node encounter rate; Variance of storage cost is negligible • Both mean and variance of storage cost and storage-time cost of time-based scheme is exponential in TTL; a priori knowledge of node encounter rate and AP encounter rate is needed to compute minimal TTL that guarantees Pd=1 What is 7DS? 7DS is a peer-to-peer data sharing network that exploits node mobility for data dissemination. Any two nodes communicate when they are in proximity. Under this transmission constraint, any pair of nodes is intermittently connected as mobility shuffles node locations. Three Categories of Applications Upload: Mobile nodes anycast data to an AP e.g. email delivery P2P: Data dissemination between mobile nodes e.g. mp3 file exchanges among iPod users Download: Content distribution of popular data e.g. movie and news video clip download Internet Real-Time LaboratoryWing Ho (Andy) YuenColumbia University • Simulation Study • Analysis compares TB and HB without feedback • Simulations compare TB and HB with and without feedback • Three scenarios: =1/1000,1/2000 and 1/5000 per sec, (17, 33 and 83 minutes to AP on average) • Node encounter rate  = 1/20 (20 sec per encounter) • For each scenario, select TB and HB schemes that has packet delivery ratio Pd=1; then consider schemes with minimum hmax for each bmax • Scenario I: (hmax,bmax)=(2,6),(3,4),(4,3),(7,2) yields Pd=1 and consumes least overhead • Email Delivery Application • A node generates an email message to be delivered to an AP. Random delay ranges from hours to days • Message delay of email delivery can be drastically reduced through packet replication at node encounters • Trade off message delay with resource consumption (storage cost and energy cost) • Duplicated packet should be purged periodically to prevent buffer overflow • Packet Purge Mechanisms • Time-based (TB): A message is purged when the TTL field expires • Hop-based (HB): Each message is appended with duple (b,h) to limit the breadth and depth of data dissemination tree • Feedback (FB): When email message is delivered to the network, 7DS may provide explicit feedback message to purge packets at all nodes • Four schemes: TB, HB, TB-FB, HB-FB • Performance Metrics • Storage cost/ Energy cost (CTB and CHB): Each packet replication incurs one unit of transmission cost at Tx node and one unit of storage cost at Rx node • Storage-time cost (STB and SHB): sum of packet storage time over all nodes. Packet storage time of node j is given by j=te-ta , where ta is the time when node j receives the packet and te is the time node j purge the packet • Quantitative Comparison of TB and HB • With no feedback, mean and variance of storage cost and storage-time cost of TB scheme is much larger than HB schemes • With feedback, mean and variance of storage cost and storage-time cost of TB scheme is on par with HB schemes • Among all HB schemes, two hop scheme often yields the least storage and storage-time costs • In practice, we can select smallest bmax such that HB(bmax,2) yields Pd=1 • Conclusion • Although 7DS does not offer delay guarantee for packet delivery, reliable and expedite data delivery is still possible, at a moderate resource overhead • All schemes under consideration have Pd=1. Packet delivery time is at the order of 100sec for all schemes in scenario I. Storage cost = 50 packets without feedback, 18 packets with feedback Time Based Xn+1=Xn+Yn-Rn E[Xn+1]=(1-+)n+1 Var[Xn]=(1+)n-1((1+)n-1) For 0 E[CTB]=(1+)TTL E[STB]=((1+)TTL+1-1)/ Var[CTB]=(1+)TTL-1((1+)TTL-1) Var[STB]=((1+)2TTL+1-(1+)TTL(2TTL+1)-1)/2 Assumptions Single email message upload Discrete time model Poisson node encounter and AP encounter Number of nodes >> number of duplicated packets Notations Xn number of duplicated packets at time n Yn number of offspring nodes receiving a packet copy at time n Zj,n number of offspring of node j at time n Rn number of nodes that arrives at an AP at time n Hop-Based For =0 E[CHB]=(bmaxhmax-1)(bmax-1) Var[CHB]=0 E[SHB]=(bmaxhmax-1)(bmax-1) bmax/ Var[SHB]=(bmaxhmax-1)(bmax-1) bmax/2 For >0 E[CHB]=(E[M]hmax+1-1)(E[M]-1) For More detailed information about this project please contact andyyuen@cs.columbia.edu.