Performance Measurement of Computer Networks

Performance Measurement of Computer Networks

Federico Montesino Pouzols, Angel Barriga Barros, Diego R. Lopez, Santiago Sánchez-Solano
Copyright: © 2008 |Pages: 7
DOI: 10.4018/978-1-59904-885-7.ch160
(Individual Chapters)
No Current Special Offers


In this article, general findings about Internet traffic models are first reviewed, with emphasis on two important invariants or characteristics that are observed with some reproducibility and independently of the precise settings of the network under consideration: self-similarity and heavy-tail marginal distributions. Then metrics and measurement techniques and tools will be discussed. This article deals with generic network performance measurement systems and outlines models, measurement techniques and tools that measure performance at the network and transport layers and can thus be applied regardless of the application layer protocols being employed. These systems are useful for analyzing performance of any network application and are an important foundational tool for enabling advanced virtual organizations (Foster, Kesselman, & Tuecke, 2001). Note, however, that application-level (or specific application details aware) measurements are commonly needed to complement generic tools so as to achieve a clear understanding of overall applications performance, which cannot be synthesized from lower level data with ease (Andrews, Cao, & McGowan, 2006).

Key Terms in this Chapter

Jitter: (Inter-packet delay variation) is the difference between the one-way-delay of two or more packets belonging to a stream of packets.

Self-Similarity: Traffic invariant commonly observed in computer networks. Common applications, such as WWW, e-mail or FTP exhibit self-similar traffic patterns. Self-similarity implies that a change of the time scale is equivalent to a change in state space scale. For discrete processes, self-similarity can be described as distributional invariance upon aggregation and scaling.

Packet Round-Trip Delay: The time for a packet to make the round trip from a source (possibly a client) to a destination (possibly a server) and back, also referred to as round-trip time (RTT). RTT can be separated into several components: forward delay, server delay and reverse delay.

Packet Delay: The time for a packet to be received. Also referred to as latency. Two types of delay are commonly measured: one-way and round-trip. For non-interactive playback operations, only the variable part of delay (jitter) is important. On the contrary, the total delay is also important for interactive services.

Packet One-Way Delay: The time for a packet to be received at a destination since it was sent from a source. Total delay can be separated into the following components: the time it takes for the source to send it, the time it takes the packet to travel along the physical links that make up the end-to-end path, the time it takes to pass through routers between those links and the time required for the server to process an incoming packet.

Throughput: The rate at which data is sent through the network, usually expressed in bits per second (bps), bytes per second (Bps) or packets per second (pps). Throughput most commonly refers to the total data transfer rate for all traffic being carried, but it can be useful to measure throughput at finer granularity, for example, for Web transactions, for Voice over IP, to specified destinations, and so forth.

Short-Range Dependence (SRD): A process is said to be short-range dependent if the dependence among the observations diminishes fast. More formally, those processes with covariance sequence that are absolutely summable are said to be short-range dependent. Most standard time-series models, such as Markov and regression models, assume short-range dependence. SRD models have a correlation structure that is significant for relatively small lags and thus cannot model network traffic in most scenarios.

Long-Range Dependence (LRD): The property of phenomena that exhibit dependence upon large time scales. More formally, LRD refers to a very slow decay of the autocovariance function. LRD or long memory processes are intimately related to self-similar processes and the two words are often used (improperly) exchangeably in the network research literature. LRD processes are asymptotically self-similar.

Capacity: Nominal physical link capacity is the theoretical maximum amount of data that a link can support, whereas IP layer link capacity is the maximum number of IP layer bits that can be transmitted over a link. Both metrics are usually measured in bits per second.

Traffic Burstiness: Network traffic is said to be bursty in the sense that a large fraction of the workload or volume of the data transferred is due to rare but large transfer volumes. In general, burstiness is dominant in data flows generated by a single user, but traffic often appears to be bursty at both single-user and multiple-user levels. More formally, a bursty time series has an empirical complementary distribution function with very slow decay. Burstiness is also referred to as impulsiveness and extreme variability.

Complete Chapter List

Search this Book: