Network Performance Measurement and Analysis

Measurement and Analysis Overview

Size, complexity and diversity of the Internet makes it very difficult to understand cause-effect relationships Measurement is necessary for understanding current system behavior and how new systems will behave How, when, where, what do we measure? Measurement is meaningless without careful analysis Analysis of data gathered from networks is quite different from work done in other disciplines Measurement/analysis enables models to be built which can be used to effectively develop and evaluate new techniques Statistical models Queuing models Simulation models

Determining What to Measure

Before any measurements can take place one must determine what to measure There are many commonly used network performance characteristics Latency Throughput Response time Arrival rate Utilization

 Bandwidth Loss Routing Reliability

Measurement Introduction
Internet measurement is done to either analyze/characterize network phenomena or to test new tools, protocols, systems, etc. Measuring Internet performance is easier said than done What does performance mean? Workload (what and where youre measuring) selection is critical Reproducibility is often essential Many tools have been developed to measure/monitor general characteristics of network performance traceroute and ping are two of the most popular These are examples of active measurement tools Passive tools are the other major category Representative and reproducible workload generation will be a focus

Active Measurement Tools

Send probe packet(s) into the network and measure a response Ping: RTT and loss Zing: one way Poisson probes Traceroute: path and RTT Nettimer (Lai): latest bottleneck bandwidth using packet pair method

 Pathchar: per-hop bandwidth, latency, loss measurement Pchar, clink: open-source reimplementation of pathchar Problem: measurement timescales vary widely

Passive Measurement Tools

Passive tools: Capture data as it passes by Logging at application level Packet capture applications (tcpdump) uses packet capture filter (bpf,libpcap) Requires access to the wire Can have many problems (adds, deletes, reordering) Flow-based measurement tools SNMP tools Routing looking glass sites Problems LOTS of data! Privacy issues Getting packet scoped in backbone of the network

Workload Generation
Local and/or wide area experiments often require representative and reproducible workloads How do we select a workload? Currently HTTP makes up the majority of Internet traffic Trace-based workloads Capture traces and replay them Black-box method Synthetic workloads Abstraction of actual operation May not capture all aspects of workload Analytic workloads Attempt to model workload precisely Very difficult

SURGE Web Workload Generator

Scalable URl Generator Analytic workload generator Based on 12 empirically derived distributions of Web browsing behaviror Explicit, parameterized models Captures heavy-tailed (highly variable) properties of Web workloads Widely used

 SURGE components: Statistical distribution generator Hyper Text Transfer Protocol (HTTP) request generator

Workload characteristics captured in SURGE

SURGE Architecture

Analyzing Measured Data

Analyzing measured data in networks is typically done using statistical methods Selecting appropriate analysis method(s) is critical Averaging Dispersion (variability) Correlations Regression analysis Distributional analysis Frequency analysis Principal-component analysis Cluster analysis Each form of analysis has strengths and weaknesses

Queuing Models
One of the key modeling techniques for computer systems in general Vast literature on queuing theory Nicely suited for network analysis Prof. Mary Vernon is our local expert Generally, queuing systems deal with a situation where jobs (of which there are many) wait in line for a resource (of which there are few) Queuing theory can enable us to determine response time Examples? Example: packets arriving at a router how can we determine how long it takes for packets to be forwarded by the router? Characteristics necessary to specify a queuing system Arrival process Service time distribution Number of servers System capacity (number of buffers) Population size Service discipline Kendal notation: A/S/m/B/K/SD Response time = waiting time + service time For stability, mean arrival rate must be less than mean service rate

Littles Law
One of the most basic theorems in queuing theory (1961) Mean number jobs in system = arrival rate * mean response time Treats a system as a black box Applies whenever number of jobs entering the system equals number of jobs leaving the system No jobs created or lost inside system Can be extended to include systems with finite buffers

Simulation Models
Simulation is one of the most common/important methods of analysis/modeling Typically an abstraction of the system under consideration Can provide significant insight to systems behavior Network simulation is difficult because of the different layers of operation and the complexity at each layer Simulation options: build your own, use someone elses Canonical network simulator is ns developed at LBL