Precision Measurements with the EVERGROW Traffic Observatory Péter Hága István Csabai

CNL - Network Performance Measurement Group 2 Outline Infrastructure of the EVERGROW Traffic Observatory Capability of the probing nodes (DAG, GPS) packet capture time synchronization Advantages of One-way measurements Advantages of Large Probing Network Study of the Properties of Large Probing Networks Tomography link delay decomposition (an example) Summary

CNL - Network Performance Measurement Group 3 Infrastructure of the Evergrow Network Probing nodes: precise time stamps with DAG accurate time synchronization with GPS effective control of co-operating nodes efficient probing technics

CNL - Network Performance Measurement Group 4 Capability of the DAG Capture Card precise timestamps with DAG 3.5E and 3.6GE time stamping the incoming and outgoing packets the accuracy of the timestamps is below 100 ns timestamps independent from the network interface handling of the probing node (without this the IRQs and the scheduling of CPU can cause serious problems) avoid measurement errors which come form the service and buffering time in the probing PC max. bandwidth to capture: 1Gb/s, with variable capture length (from the header only to all the packet), resulting a tcpdump like output ability of the time synchronization (independently from NTP) packet capture and scheduled send (with a new model of DAG card)

CNL - Network Performance Measurement Group 5 Time Synchronization with GPS DAG calculates time from: external reference source (NTP server) to reach the precise second value (only at initialization) Garmin 35 HVS cost effective GPS receiver + self-developed converter card GPS One-pulse-per-second (PPS) signal to calculate the correct frequency of the clock PPS is a periodic (1Hz) synchronization signal with high accuracy (100 ns)

CNL - Network Performance Measurement Group 6 Co-operation of Nodes & Efficient Probing Technics Other important advantages of EVERGROW Traffic Observatory: co-operation between nodes ability of the usage of high tech methods and tools wide variety of monitoring and probing methods not only Ping and Traceroute Importance of the time synchronization: with these methods the clock of the DAG can run without skew without time synchronization the skew of the clocks in the nodes makes the results of large number of measurements incomparable

CNL - Network Performance Measurement Group 7 Advantages of One-way measurements Measured data contains: only the effects of the investigated path nor the effects of the backward path neither other service time (e.g. ICMP return) high time resolution: accurate delay values for Packet Pair, Packet Quartet methods correlated packet sending to measure Y-topology With time synchronization: comparing and using the results of measurements between different e2e pairs for the same estimation long term stability

CNL - Network Performance Measurement Group 8 Advantages of Large Probing Network It can be useful to make: Europe wide network monitoring (delays, loss, available bandwidth, trends, performance analysis, etc.) development, investigation and verification of new methods and applications on many real path tomography (it can decrease the invasive property of many active probing methods) investigations on interactions between many probing streams probe sending with controlled correlation effect of controlled cross traffic on controlled probe stream

CNL - Network Performance Measurement Group 9 Studying the Properties of a Large Network Preliminary study of large networks: self-developed topology generator (TopoGen) NLANR 1 data collection (US academic backbone traces) investigation of the ratio of the number of internal nodes and probing nodes in a specific, real network topology ratio of the number of links and measurable paths (between two probing nodes) investigation of the role of placing nodes with different configuration 1: Supported by the National Science Foundation Cooperative agreement nos. ANI (2002) and ANI (1998), and the National Laboratory for Applied Network Research.

CNL - Network Performance Measurement Group 10 Studying the Properties of Large Networks Topological properties: Number of internal nodes vs. number of EVERGROW nodes Number of internal links Number of irreducible links along the path between the nodes importance of the placing of the probing nodes

CNL - Network Performance Measurement Group 11 Tomography The results form all the e2e pairs can be extended: estimating correlations between the overlapping paths can give details about the overlapping and non-overlapping part of the path inverse problem: for sum of additive parameters (delay) the values on internal links can be calculated mimicking multicast probing can be also useful to inference link-level parameters it needs quick & accurate measurements

CNL - Network Performance Measurement Group 12 Example for the usage of Tomography The studied problem: determine the individual delays of all the links in our network 1st solution: Traceroute like methods: one measurement for each link like this we need LN 2 measurements to collect data for this we need L 3 N 2 time (network conditions can change during this) 2nd solution: with tomography and SVD: one measurement for each path for this we need N 2 measurements and LN 2 time typical value of L is about 13, the whole time of the measurement can be only 0.6% with tomography, with only 0.6% of the probing packets, compared to the traceroute like measurements

CNL - Network Performance Measurement Group 13 How does it work? additive property of the links is the propagation delay discovering the topology, and simplifying all the reducible path to the simplest graph decompose the delays of paths to the individual links delay with SVD SVD can work with over (for large networks) and under (for small networks) determined cases, too # links ~ # probing nodes, # paths ~ (# probing nodes) 2 result: with only one measurement between the end points we could calculate almost all the link delays separately Example for the usage of Tomography

CNL - Network Performance Measurement Group 14 Example for the usage of Tomography (A)(l)=(b) SVD resulting the individual link delays

CNL - Network Performance Measurement Group 15 Example for the usage of Tomography Goodness of the estimation: without noise in the measurement noise makes the equations much or less inconsistent noise with different amplitude noise with different distributions Delay decomposition in Random graph

CNL - Network Performance Measurement Group 16 Example for the usage of Tomography Eliminate the effect of the noise: averaging the results of the measurements filtering (e.g. for minimum values at delays)

CNL - Network Performance Measurement Group 17 TopoGen, NLANR, EVERGROW Different topologies: Random Graph (Erdős-Rényi) Small World (Strogatz-Watts) Scale Free (Albert-Barabási) NLANR (to study the effect of real noise) noisy, highly inconsistent path delays Evergrow Traffic Observatory...

CNL - Network Performance Measurement Group 18 Summary the EVERGROW Traffic Observatory will provide: large number of probing nodes and paths to investigate precise time stamping at the nodes accurate clock, synchronized probing nodes e.g. quick (with small noise), accurate delay measurements efficient methods and data collection monitoring several network properties testbed for different new methods or applications with real background traffic and real topology