Title Introduction Methodology Experimental Results Conclusion Backup BitTorrent Experiments on Testbeds: A Study of the Impact of Network Latencies Ashwin Rao , Arnaud Legout, and Walid Dabbous INRIA, Projet Plan` ete (ashwin.rao,arnaud.legout,walid.dabbous)@inria.fr 1 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Outline Introduction Methodology Experimental Results Homogeneous Latency Heterogeneous Latency Conclusion 2 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Overview of BitTorrent 3 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Evaluation of BitTorrent Performance Grid5000 sites Planetlab sites [planete-lab.org] [grid5000.fr] • Grid5000 and PlaneteLab Testbeds 4 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Evaluation of BitTorrent Performance Grid5000 sites Planetlab sites [planete-lab.org] [grid5000.fr] • Grid5000 and PlaneteLab Testbeds • Absence of network latency between • Instances of application running on same machine • Machines in the same LAN 4 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Evaluation of BitTorrent Performance Grid5000 sites Planetlab sites [planete-lab.org] [grid5000.fr] • Grid5000 and PlaneteLab Testbeds • Absence of network latency between • Instances of application running on same machine • Machines in the same LAN Does network latency affect the outcome of BitTorrent experiments performed on testbeds 4 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Can Latency Impact BitTorrent Performance? TCP Ramp-up and Impact of Upload Rates 5 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Can Latency Impact BitTorrent Performance? Delays in receiving BitTorrent control TCP Ramp-up and Impact of messages Upload Rates 5 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Can Latency Impact BitTorrent Performance? Delays in receiving BitTorrent control TCP Ramp-up and Impact of messages Upload Rates Can testbeds such as Grid5000 be used for experimental evaluation of BitTorrent? 5 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Outline Introduction Methodology Experimental Results Homogeneous Latency Heterogeneous Latency Conclusion 6 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Testbed Setup 4 Machines of Grid5000 Experimental Testbed 7 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Testbed Setup Machines capable of running 100 instances of BitTorrent Client 7 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Testbed Setup Peers on same machine communicate using Loopback device 7 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Testbed Setup Peers on different machines communicate using Ethernet device 7 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Testbed Setup Tracker and Seed placed on same machine 7 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Experiment Setup • Experiment Parameters • 50 MB file • 1 Tracker, 1 Seed, and 300 Leechers • Upload rates - 10 KiB/s to 100 KiB/s • Emulate RTT - 1000 ms • Metric • Download completion time 8 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Outline Introduction Methodology Experimental Results Homogeneous Latency Heterogeneous Latency Conclusion 9 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Setup Same delay on Loopback and Ethernet Device 10 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Download Completion Time vs RTT 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results 1000 ms RTT does not increase Average Download Completion Time by more than 15% 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Download Completion Time not a monotonously increasing function of RTT 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Download Completion Time not a monotonously increasing function of RTT 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Impact of latency not observed when seed is fast 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Marginal impact of RTT 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Homogeneous Latency - Experiment Results Marginal impact of RTT 11 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Heterogeneous Latency Each machine used to abstract an AS 12 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Heterogeneous Latency Different latency on loopback and ethernet device 12 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Results - Small RTT Between Peers RTT between a pair of Upload Rate limited to leechers. 20 kB/s. RTT between any two peers is less than 100 ms 13 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Results - Small RTT Between Peers RTT between a pair of Upload Rate limited to leechers. 20 kB/s. RTT between any two peers is less than 100 ms Marginal Impact of Latency 13 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Results - Small RTT Between Peers RTT between a pair of Upload Rate limited to leechers. 20 kB/s. RTT between any two peers is less than 100 ms Marginal Impact of Latency 13 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Results - Large RTT Between Some Peers RTT between a pair of Upload Rate limited to leechers. 20 kB/s. RTT between some of the peers is greater than 200 ms 14 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Results - Large RTT Between Some Peers Upload Rate limited to RTT between a pair of 20 kB/s. leechers. RTT between some of the peers is greater than 200 ms Marginal Impact of Latency 14 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Outline Introduction Methodology Experimental Results Homogeneous Latency Heterogeneous Latency Conclusion 15 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Conclusion • Latency has a marginal impact on the outcome of BitTorrent experiments 16 / 17
Title Introduction Methodology Experimental Results Conclusion Backup Conclusion • Latency has a marginal impact on the outcome of BitTorrent experiments BitTorrent experiments can be performed on testbeds without explicitly emulating latency. 16 / 17
Title Introduction Methodology Experimental Results Conclusion Backup BitTorrent Experiments on Testbeds: A Study of the Impact of Network Latencies Ashwin Rao (ashwin.rao@inria.fr) 17 / 17
Title Introduction Methodology Experimental Results Conclusion Backup TCP Segmentation Offloading 1 0.8 TSO Enabled 0.6 0 ms 0.4 100 ms 0.2 400 ms 0 1 10 100 1000 10000 1 0.8 WIDE backbone CDF 0.6 0.4 0.2 WIDE backbone 0 1 10 100 1000 10000 1 0.8 TSO Disabled 0.6 0 ms 0.4 100 ms 0.2 400 ms 0 1 10 100 1000 10000 TCP Payload Length (bytes) Figure: Impact of Segmentation Offloading − / 16
Title Introduction Methodology Experimental Results Conclusion Backup Number of Nodes on a Machine 1100 1000 RTT estimate of TCP (ms) 900 Minimum RTT 0 ms 800 Minimum RTT 400 ms Minimum RTT 1000 ms 700 600 500 400 300 200 100 0 0 25 50 75 100 Number of leechers running on a given machine Figure: Impact of Number of Nodes on a Machine Machines support up to 100 instances of a leecher − / 16
Recommend
More recommend