Passive Calibration of an Active Measurement System
Stephen Donnelly*, Ian Graham*, Rene Wilhelm**
* The University of Waikato
** RIPE NCC
Measuring QoS in the internet today is difficult as notions of what
constitutes QoS vary. Service Level Agreements between customers and
network service providers are often poorly defined. The most workable
SLAs currently are defined in terms of how long packets may take to
travel from some host to another host, and what acceptable levels of
packet loss are.
This paper presents a direct comparison of active One-way-Delay
measurements with an independent passive measurement system over an
intercontinental distance.
The IETF's IP Performance Metrics Working Group have codified metrics
for One-way-Delay[1], One-way-Loss[2], and Round-trip-Delay[3], which
should be useful to people writing SLAs. The IPPM WG also discuss
practical methods for making measurements of their metrics[4]. This
involves actively sending a periodic or poisson stream of test packets
from one host to another, and observing the arrivals at their
destination.
There are at least two implementations of measurement systems to collect
these metrics. One is by ANS[5], and the other by RIPE NCC[6].
Since 1997, RIPE NCC has been running an experimental system called
'Test Traffic Measurement', which collects Type-P-One-way-Delay,
Type-P-One-way-Loss, and routing vectors across a full mesh of their
'testboxes'. There are currently approximately 50 testboxes distributed
mostly amongst memeber ISPs, with a few outside Europe, including the
USA, and one in New Zealand.
The testbox is a PC running FreeBSD, with a conventional Ethernet NIC.
Each testbox also contains a GPS receiver, which is used for
timekeeping. This global source of synchronised timing information
allows the measurement of one way metrics, by placing timestamps inside
packets at the source, and sending them to a destination machine. At the
destination the timestamps can be compared to the synchronised clock.
This approach should lead to a very accurate and stable measurement
platform. It is however quite difficult to characterise the error
distribution of such a system over a large distance. Previous work[7]
compared the RIPE NCC system with Surveyor, but as both are active
systems, and do not necessarily probe the network at the same time, only
a general agreement between the systems could be inferred.
The WAND group DAG monitor[8] is a hardware and software system that
passively receives and timestamps network traffic in hardware. The DAG
includes a hardware clock that is used for timestamping which is
constantly and smoothly adjusted using a GPS reference to within half a
microsecond of UTC.
Two of these systems at two RIPE NCC testbox sites have passively
collected and timestamped the testbox active probes, providing an
independent measure the One-way-Delay of the packets from the time they
leave one testbox until they arrive at their destination. The passive
measurements at each site provide a very accurate measure of the
difference between the testbox source timestamp, and the time at which
the packet actually appears on the source network. This time difference
is referred to as the transmission latency. The difference between the
packet arriving at the destination network and the testboxes timestamp
of the event can be similarly found. This is referred to as reception
latency, and quantifies the end host packet reception behaviour. Finally
the passive timestamps are compared to produce a 'wire One-way-Delay'
measurement which is compared to the corresponding testbox One-way-Delay
result.
A high level of agreement was found between the active and passive
systems in this study. An explanation of the distribution of differences
between the systems results is offered.
It is hoped that these results will help researchers making active
measurements to understand error contributions from end-point equipment
when interpreting their results.
- "A One-way Delay Metric for IPPM" G. Almes, S. Kalidindi, M.
Zekauskas RFC2679, September 1999.
- "A One-way Packet Loss Metric for IPPM" G. Almes, S. Kalidindi, M.
Zekauskas RFC2680, September 1999.
- "A Round-trip Delay Metric for IPPM" G. Almes, S. Kalidindi, M.
Zekauskas RFC2681, September 1999.
- "Network performance measurement for periodic streams" V. Raisanen,
G. Grotefeld Internet Draft , July 2000.
- "Surveyor: An Infrastructure for Internet Performance Measurements"
Sunil Kalidindi, Matthew J. Zekauskas ANS Proc. INET'99
, June
1999.
- "Internet Delay Measurements using Test Traffic" Henk Uijterwaal,
Olaf Kolkman RIPE NCC Document: RIPE-158, May 1997.
- "Comparing two Implementations of the IETF IPPM One-way Delay and
Loss Metrics" Sunil Kalidindi, Matt Zekauskas and Henk Uijterwaal Proc.
PAM2000, April 2000.
- "Nonintrusive and Accurate Measurement of Unidirectional Delay and
Delay Variation on the Internet" I. D. Graham, S. F. Donnelly, S.
Martin, J. Martens, and J. G. Cleary Proc. INET'98, July 1998.
|
