Side_1_360

length distributions and link utilization. Hence,

these metrics can be computed locally at the PC

that captures the packet trace.

3.4.3 Multiple Packet Traces

In order to compute unidirectional network level

performance metrics like delay and loss, it is

necessary to correlate the information in multi-

ple packet traces. Thus, information from the

packet traces collected at various measurement

points must at regular intervals be exported to

a central host for post-processing.

To recognize the same packet in several packet

traces it is necessary to use a combination of IP

and transport protocol header fields like source

and destination IP address, source and destina-

tion port number, and identification field

[Fraleigh]. The delay of a packet from one mea-

surement point to another measurement point is

then determined from the associated timestamps.

A packet seen at an ingress measurement point

that is not observed at its corresponding egress

measurement point within a given time-out inter-

val is defined as lost.

3.5 Handling and Post-processing of

Packet Traces

This section compares various approaches to

post-processing of packet traces in context of the

scenario shown in Figure 3.4. A measurement

unit capable of capturing packet traces is located

at every ingress and egress link of the network

domain.

Figure 3.5 shows the various steps in the han-

dling and post-processing of measurement data.

The raw packet trace captured at a measurement

point can be grouped and sorted according to

various schemes. Examples of ways to group

and sort the packets contained in a packet trace

include the following:

• No grouping or sorting of packets;


• Group packets according to end-to-end flows
  [Claffy95];


• Group packets that follow a certain end-to-end
  path;


• Group packets by incoming and outgoing
  interface of a node;


• Group packets by the application that gener-
  ated the packet.

The next step in the post-processing of measure-

ment data is to reduce the volume of the mea-

surement data. Data reduction techniques can be

divided into two classes:

Figure 3.3 Passive measurements

Figure 3.4 Data flow and storage for passive measurements

Network section

Single trace

• throughput


• utilization

Single trace

• throughput


• utilization

Multiple trace

• unidirectional delay


• unidirectional loss

Central

processing

Correlate

information

Network-level Passive measurement properties

metric

Delay + Measure performance as experienced by real packets.

Loss + Do not disturb the operation of the network.

• Resource intensive (huge data volumes). This can be
  handled by using data reduction techniques.

Throughput + Measure performance as experienced by real packets.

Utilization + Do not disturb the operation of the network.

+ Estimated from a single packet trace.

Table 3.2 Passive measurements of performance metrics

Data

Data

Data

Central host

λi

Monitor i

Data

Data