1、PlanetLab Deployment and Analysis of Network Coordinate Systems,Fenglin Liao Keshava Subramanya Veljko Pejovicfenglin,keshava,veljko cs.ucsb.edu,Contents,Motivation for the Project Project Objectives Experimental Setup Challenges Results Conclusions Guidelines for the future Credits,Motivation for t
2、he project,Non-availability of exhaustive NC system comparison Need for relevant metric for comparing NC systems Need for neutral analysis of NC systems under various criteria *,A tale of 2 NC Systems,Vivaldi: Simple, adaptive, distributed, lightweight GNP: Centralized (uses landmarks), static,Proje
3、ct Objectives,Accuracy RTT vs. Calculated distances Effect of node “distance” Convergence time Robustness Effect of network traffic Impact of malicious nodes Stability Behavior of system when nodes are joining and leaving (Vivaldi) Impact of Landmarks leaving the system (GNP),Experimental Setup,Sele
4、cted 100+ on PlanetLabs Sorted and selected from every class based on RTT Ran Pyixda - a Java implementation of Vivaldi based closely on original MIT Paper (Azureus) GNP implementation by T. S. Eugene Ng, author of the original paper Constructed full ping mesh with periodic ping measurements Modifie
5、d code of Pyxida, created daemons, mined log file data using scripts, analyzed data using Matlab Total run time of over 150+ hours,Challenges,Most importantly, being comprehensive, relevant Bringing dissimilar NC systems to comparable platform Large amounts of data to be processed Need to move massi
6、ve amounts of Data to PL Killed some csil machines in the process Angry emails from local sys-admins due to “ssh probe traffic”,Results - Accuracy,GNP, as expected, has better accuracy Increasing no. of dimensions of Vivaldi improves performance but does not beat GNP 65% of Vivaldi, and 75% of GNP m
7、easurements have relative error less than 0.2, hence, fairly accurate!,Results - Accuracy: RRL and CNL,RRL: Relative Rank Loss how well a network coordinate scheme respects the relative ordering of all pairs of neighbors (Lua et al) CNL: Closest Neighbor Loss Percentage of nodes whose estimated clos
8、est neighbor differs from the actual one (Lua et al) Provides a NC independent platform for analysis of accuracy Most applications need just these two, simplifies API set,RRL: Vivaldi vs. GNP,Lower RRL in Vivaldi Lower CNL with GNPSpikes in GNP for some nodes,Results - Accuracy: Effect of Distance,V
9、ivaldi overestimates closer node distances GNP underestimates far away nodes,Results - Accuracy: Convergence time,Most nodes converged (err0.2) within 2 hours for Vivaldi GNP Static (calculation was CPU intensive),Results - Robustness: Effect of Network traffic,Selected 8 nodes and ran Vivaldi and G
10、NP Subjected the network to heavy traffic Compared measurements before, during and after trafficResult: No significant changes for Vivaldi and GNP Possibly because both are really light weight on the network,Results - Robustness: Impact of Malicious Nodes,Set up a stable ring of 8 nodes Hacked the V
11、ivaldi code to send out constant “GossipMessages” with very large values of coordinates Added 3 malicious nodes to the network,Malicious Nodes injected at T=2.5h,Vivaldi blown away by the attack GNP Malicious Landmark beyond the scope of this work, but Landmarks are usually trusted nodes,Results - S
12、tability: Effect of Network Churn,Vivaldi reacts more to node churning, but recovers soon Nodes leaving GNP NC have no effect on other nodes Landmarks leaving the network has almost no effect on GNP stability “Back-up” landmarks are often available to ensure system stability during landmark failure
13、(called candidate landmarks) If the number of available landmarks go below minimum threshold, NC system fails to continue operation,CDF,NC System Confidence factor,3 Nodes Left,GNP Relative Error 9 to 15 Landmarks,3 Nodes Joined,Conclusions,GNP is generally more accurate but both are still not accur
14、ate enough CNL performance of both systems are highly inaccurate - more than 50% error GNP makes really high RRL on some nodes, hence not very useful for applications relying on RRL Based on distance based measurements, Vivaldi is suited for large networks and GNP for smaller and denser networks Viv
15、aldi takes about 2.5 hrs to converge. Using it on high churn networks may not be ideal,Guidelines for the future,A hybrid approach: Vivaldi like, but with periodic random ping to improve accuracy Make light weight (a C implementation) Multiple implementations (two running - choose one or average est
16、imated RTT) Log files to “remember” coordinates of last run (brings down convergence time) Vivaldi should guard itself against malicious nodes by using some reputation mechanisms Make mobile/wireless nodes “read-only”,Thank You,Questions/Comments?,CreditsLedlie, Pietzuch, Parker, http:/ S. E. Ng and H. Zhang. Predicting Internet network distance with coordinates-based approachesF. Dabek, R. Cox, F. Kaashoek, and R. Morris. Vivaldi: A Decentralized Network Coordinate System. In SIGCOMM, August 2004Network Traffic Generator: http:/ Alagud (UCSB) For his Java expertise and cheerful helping hand,
