generated by specialized hardware. This solution
offers accuracy in the range of a few microsec-
onds [Pasztor] [Fraleigh].3.3 Active Measurements
Active measurements at the network-level are
carried out by inserting probe packets and ob-
serving these probe packets. The major assump-
tion behind this approach is that the performance
of probe packets is representative of the perfor-
mance experienced by real packets. To measure
the performance of a service, the traffic charac-
teristics should be considered (interarrival time
of packets, packet lengths, etc.). Note that active
measurements are intrusive and therefore the
measurements affect the system being measured.Measurements of round-trip times and packet
loss in IP networks are usually performed by
software based on the ICMP Echo Reply/
Request messages [RFC792] (ping) running
on PCs. The advantage of this approach is that
most implementations of TCP/IP support ICMP
Echo messages. Thus, measurements can be per-
formed by pings to almost any host in the net-
work without making any special arrangements
beforehand. This approach only requires the
remote host to respond to ICMP “ECHO
request” messages [RFC792]. Since ICMP uses
IP services, it measures network-level perfor-
mance rather than transport level performance.
The major drawback is that ICMP Echo Reply/
Request measurements normally are limited to
measure bi-directional delays. Other disadvan-
tages by using ping are that the routers along the
path may treat ICMP packets differently from
other IP packets, block or limit the rate of ICMP
packets (Firewalls, etc.).3.3.1 Active Measurements of Unidirec-
tional Network-level Performance
Parameters
The Surveyor and Ripe test traffic projects mea-
sure unidirectional performance by injecting
probe packets. The dedicated measurement PCs
are synchronized by using GPS receivers. The
general principle is as follows; to measure the
performance between a given source and desti-
nation pair, the source injects probe packets
addressed to the destination. The sender adds a
timestamp and sequence number to every probe
packet. That is, “one-way pinging” and through-
put tests are performed using dedicated measure-
ment units located at selected measurement
points. The injected probe packets can be either
in-service4)or out-of-service. That is, one can
either insert dedicated test flows or insert test
packets inside user flows. [Lindh] proposes a
measurement infrastructure embedded in net-
work nodes that is based on inserting special
Operations and Maintenance (OAM) packets
into the user traffic. This method is similar to the
OAM cells used in ATM networks [Prycker95].
The measurement units must have capabilities
to post-process, store, and export the collected
measurement data.Table 3.1 shows the fundamental network-level
performance metrics and the suitability of the
active method. It may be noted that unlike pas-
sive measurements the active measurements can-
not collect detailed information about properties
of user generated packets, e.g. traffic mixture,
packet length and network traffic matrix.3.4 Passive Measurements
Passive measurement data is collected by ob-
serving real packets at selected measurement4)The method requires further study in the context of IP networks.
Network-level metric Characterization of active measurementsUnidirectional delay + Straightforward to implement.
Unidirectional packet + Easy to aggregate performance metrics at a single measurement
loss unit.- Are the measurements representative of the performance real
packets experience? - What should the traffic pattern of probe packets be? What packet
interarrival time and packet lengths are representative for various
service classes? - The injected probe packets disturb real traffic (Heissenberg bug).
Throughput – How can representative measurements of throughput/utilization
Link utilization be performed?- Active measurements generally disturb the operation of the network.
Thus, lightweight tests are needed.
Table 3.1 Active measure-
ments of performance metrics