TCP Spatial-Temporal Measurement and Analysis

Similar documents
Spatial-Temporal Analysis of passive TCP Measurements

The War Between Mice and Elephants

CS4700/CS5700 Fundamentals of Computer Networks

On Network Dimensioning Approach for the Internet

Localization of network performance problems with multi-level discrete tomography

Christos Papadopoulos

DOMINO: A System to Detect Greedy Behavior in IEEE Hotspots

Introduction to Networking and Systems Measurements

Basics (cont.) Characteristics of data communication technologies OSI-Model

Passive Packet Loss Detection in Wi-Fi Networks and its Effect on HTTP Traffic Characteristics

XORs in the Air: Practical Wireless Network Coding

Available Bandwidth Estimation. Probing Packet Train in Pathneck. Transmission of RPT. Choke Point Detection. Packet train probing

ECE 598HH: Special Topics in Wireless Networks and Mobile Systems

Network Measurement (II) Network Tomography

AS Router Connectedness Based on Multiple Vantage Points and the Resulting Topologies

Fast Retransmit. Problem: coarsegrain. timeouts lead to idle periods Fast retransmit: use duplicate ACKs to trigger retransmission

Improving Accuracy in End-to-end Packet Loss Measurement

Congestion Collapse in the 1980s

Efficient solutions for the monitoring of the Internet

Mobile Routing : Computer Networking. Overview. How to Handle Mobile Nodes? Mobile IP Ad-hoc network routing Assigned reading

Improving the Robustness of TCP to Non-Congestion Events

Unicast Probing to Estimate Shared Loss Rate

The Bootstrap and Jackknife

Congestion Control In The Internet Part 2: How it is implemented in TCP. JY Le Boudec 2015

CS419: Computer Networks. Lecture 10, Part 3: Apr 13, 2005 Transport: TCP performance

Lecture 13: Traffic Engineering

Investigating the Use of Synchronized Clocks in TCP Congestion Control

Using Concurrent Multipath Transmission for Transport Virtualization: Analyzing Path Selection

Distributed Anomaly Detection using Autoencoder Neural Networks in WSN for IoT

Quantifying Internet End-to-End Route Similarity

End-to-End Mechanisms for QoS Support in Wireless Networks

Lecture 5: Performance Analysis I

Call Admission Control in IP networks with QoS support

Diagnosis of TCP Overlay Connection Failures using Bayesian Networks

Rate Adaptation in

Transport Layer (Congestion Control)

PoP Level Mapping And Peering Deals

Internet Measurements. Motivation

10-701/15-781, Fall 2006, Final

Lecture (08, 09) Routing in Switched Networks

Internetworking Protocols and Software (COMP416)

AS Connectedness Based on Multiple Vantage Points and the Resulting Topologies

Clustering the Internet Topology at the AS-level

SELECTION OF METRICS (CONT) Gaia Maselli

Lecture 15: Measurement Studies on Internet Routing

Tuning RED for Web Traffic

Improving TCP Performance over Wireless Networks using Loss Predictors

The Effects of Asymmetry on TCP Performance

Energy-Aware Routing in Wireless Ad-hoc Networks

PfR Voice Traffic Optimization Using Active Probes

SCONE: A Tool to Estimate Shared Congestion Among Internet Paths

CS 344/444 Computer Network Fundamentals Final Exam Solutions Spring 2007

CHAPTER 5 PROPAGATION DELAY

Fast, Efficient, and Robust Multicast in Wireless Mesh Networks

BANDWIDTH MEASUREMENT IN WIRELESS NETWORKS

Internet trafic monitoring

TCP Nice: A Mechanism for Background Transfers

Impact of transmission errors on TCP performance. Outline. Random Errors

ECE 697J Advanced Topics in Computer Networks

DETECTING RESOLVERS AT.NZ. Jing Qiao, Sebastian Castro DNS-OARC 29 Amsterdam, October 2018

Chapter 12. Routing and Routing Protocols 12-1

Detecting Shared Congestion of Flows Via End-to-end Measurement

Supporting Service Differentiation for Real-Time and Best-Effort Traffic in Stateless Wireless Ad-Hoc Networks (SWAN)

Congestion Control In The Internet Part 2: How it is implemented in TCP. JY Le Boudec 2015

CS533 Modeling and Performance Evaluation of Network and Computer Systems

CS533 Modeling and Performance Evaluation of Network and Computer Systems

Computer Networking. Reliable Transport. Reliable Transport. Principles of reliable data transfer. Reliable data transfer. Elements of Procedure

Toward a Reliable Data Transport Architecture for Optical Burst-Switched Networks

An overview on Internet Measurement Methodologies, Techniques and Tools

Measurement Study of Lowbitrate Internet Video Streaming

Congestion Control in Datacenters. Ahmed Saeed

Intelligent Routing Platform

University of Virginia Department of Computer Science. CS 4501: Information Retrieval Fall 2015

CCNA R&S: Introduction to Networks. Chapter 7: The Transport Layer

Sherlock Diagnosing Problems in the Enterprise

Topology Discovery on Unicast Networks: A Hierarchical Approach Based on End-to-End Measurements

COMPUTER NETWORKS PERFORMANCE. Gaia Maselli

BLM6196 COMPUTER NETWORKS AND COMMUNICATION PROTOCOLS

DDoS Attacks Classification using Numeric Attribute-based Gaussian Naive Bayes

under grant CNS This work was supported in part by the National Science Foundation (NSF)

Lecture 7: Flow Control"

3. Data Preprocessing. 3.1 Introduction

The War Between Mice and Elephants

2. Data Preprocessing

Multiwavelet based-approach to detect shared congestion in computer networks

Recap. TCP connection setup/teardown Sliding window, flow control Retransmission timeouts Fairness, max-min fairness AIMD achieves max-min fairness

Part 2. Simulation Analysis

CS 260: Seminar in Computer Science: Multimedia Networking

Outdoor Wi Fi Mesh Routing Protocol. Akshata Danivasa Alphonse Hansel Anthony Mani Pandian Vikas Iyer Vinesh Pallen

Lecture 9. Quality of Service in ad hoc wireless networks

CSE 123: Computer Networks Alex C. Snoeren. HW 1 due NOW!

Isolating Compromised Routers. Alper Mizrak, Keith Marzullo and Stefan Savage UC San Diego Department of Computer Science and Engineering

Congestion Control In The Internet Part 2: How it is implemented in TCP. JY Le Boudec 2014

Latent Space Model for Road Networks to Predict Time-Varying Traffic. Presented by: Rob Fitzgerald Spring 2017

Multicasting in ad hoc networks: Energy efficient

Performance Comparison of TFRC and TCP

Telecommunication Services Engineering Lab. Roch H. Glitho

CE693: Adv. Computer Networking

Comparing TCP Congestion Control Algorithms Based on Passively Collected Packet Traces

Reinforcement learning algorithms for non-stationary environments. Devika Subramanian Rice University

Transcription:

TCP Spatial-Temporal Measurement and Analysis Infocom 05 E. Brosh Distributed Network Analysis (DNA) Group Columbia University G. Lubetzky-Sharon, Y. Shavitt Tel-Aviv University 1

Overview Inferring network performance Method of identifying root cause (bad) links Analytic and Experimental Results Inferring path characteristics from end-to-end measurements 2

Goal Infer the performance of the internet by passively monitoring network traffic Quantitative versus qualitative approach How bad is a link? => infer the characteristics (e.g., lossiness) of network links Which links are bad? => classify links according to their performance 3

Solutions - Previous Work Focus on the quantitative approach Inferring link characteristics (e.g., lossiness) Active probing (MINC, Pathchar, Packet Stripes, ) Network overhead may bias the results Passive observations (padmanabhan 03) High-complexity analysis, e.g., Bayesian inference Detecting shared congestion (passive approach) Entropy-based clustering technique (Katabi 01) Measurement point needs to observe large portions of data Correlation-based technique (Rubenstein 00) Supports only pair-wise congestion detection 4

Our contributions Focus on the qualitative approach Consider a simplified problem: identifying lowperformance links rather than inferring detailed link statistics Cost-effective algorithm Based on a Root Cause criterion Practical and relatively easy to implement High detection ratio with low errors 5

Passive Measurements Experimental Setup Capture TCP headers using tcpdump from an edge ISP router Network Topology Collect network paths using traceroute (infrequent probing) Diagnostic node ISP Internet 6

Our approach link classification Goal locate low-performance links Our approach: classify and report lowperformance links using a Root Cause criterion Performance metrics: link loss, reordering, and duplication rates Rest of the talk will focus on link loss 7

Root Cause Criterion Def: link (u,v) is a Root Cause (RC) If its loss probability is larger by at least δ compared to the loss probabilities of the links entering u or leaving v a b c 0.1 e 0.2 0.3 f 0.2 0.5 0.3>0.1 and 0.3>0.2 => c is RC link 0.5>0.3 => f is RC link δ=0 (zero threshold) 8

Root Cause Detection Algorithm 1. Estimate link loss rates For each link average the loss rates of the paths going through it 2. Link classification For each link apply the root cause criterion using the estimated link loss rates Assumption: network routes and link characteristics are stationary 9

Evaluation - Analysis Generalization: assume weighted average {w 1,,w n } = weights of paths going through link l {r 1, r n } = loss rates of the paths going through link l ^ p l = w j r j, the weighted average of losses on link l is a biased estimator pˆ l = p (1 { l + pl w j e j 123 { ) real real success prob. loss prob. j: l t error factor e j = loss probability of path j excluding l s losses j 10

Estimator Analysis Assumptions Loss probabilities are i.i.d random vars, mean 1-μ, var σ 2 All paths have length h Bias properties E[ pˆ V[ pˆ l l p p l l ] = (1 p ] = (1 p l l )(1 μ ) 2 2 (( σ h 1 2 + μ ) h 1 μ 2( h 1) Bias tends to a constant value for small loss prob. The variance of the bias can decrease as path length increases ) ) w 2 j 11

Estimator Analysis (cont d) Root Cause criterion The difference between loss probability estimators for two adjacent links has a lower bias than the bias of a single link estimator. 12

Evaluation - Simulations Methodology A directed acyclic graph topology 100-3000 nodes, max degree = 5-10 Link loss probabilities drawn from Zipf and uniform distributions over the range [0,0.04] For each loss probability value: Portion of experiments with true classification Portion of experiments with false classification (false positive and false negative) 13

Simulation Results For high loss probabilities Detection ratio above 95% False ratio below 10% Positive correlation between RCclassified links and links with true high loss probabilities. δ=0 (zero threshold) 14

Evaluation - Internet Results Israeli ISP sample, Dec. 2002, 230 million TCP packets Validation by graph visualization of problematic links Worst lossy links, highest->lowest Internal Link in a US provider Israeli ISP -- UK Software company Israeli ISP -- Israeli portal Israeli ISP -- Israeli portal Internal link in an Israeli portal Link between 2 Israeli ISPs Israeli ISP -- US telecom Israeli ISP and a US ISP 2 US ISPs Top worst links (significant losses or reordering) are inter-isp links. 15

Inferring path characteristics from end-to-end measurements Classify out-of-sequence TCP packets: Sender s retransmissions ( = loss rate) Network reordering Network duplication Our technique relies on observing a oneway direction of a TCP connection Key Idea: Leverage IP-ID field In practice the sender s IP-ID forms a monotonic increasing sequence 16

Classification Process (per each packet) Is TCP sequence # Out-of-order? yes yes Is packet previously observed? no Is the IP-ID of both Packets different? no Duplicate yes yes Retransmission Is IP-ID In-order? no Reordering 17

Summary Passive inference of low-performance links is feasible Based on a Root-Cause criterion Cost effective and comprehensive solution High detection ratio (95%) with low errors (10%) Temporal results (details in paper) Passive study of consecutive packet losses Geometric-like distribution Revisited classical TCP throughput analysis Show that TCP throughput formulas apply for the Bernoulli loss model. 18