FYI - we have been looking at the impact of packet loss and jitter on Voice
over IP perceptual quality and have found it desirable to take a different
approach. We use a 4-state Markov model to represent packet loss
distribution and update this using packet loss events (i.e. gap lengths).
This allows us to capture information on the degree of burstiness and on the
loss densities within bursts and gaps. You also avoid the issue described
by Tim in that loss densities are not related to arbitrary time periods.
Alan Clark
Telchemy
-----Original Message-----
From: owner-sls@ist-tequila.org [mailto:owner-sls@ist-tequila.org]On
Behalf Of Timothy SOETENS
Sent: Friday, July 20, 2001 8:13 AM
To: sls@ist-tequila.org
Subject: [tequila/sls] Statistical bounds and time intervals
Hi all,
Some remarks on section 3.5 of draft-tequila-sls-01.txt:
It indicates as possible guarantees:
- delay, time interval, optional quantile
- jitter, time interval, optional quantile
- packet loss, time interval
- throughput, time interval
IMHO time interval should also be an optional parameter. It is no use
indicating a time interval (granularity) if we use deterministic bounds
on delay and jitter. If the transfer delay from ingress to egress
measured in any time period of x minutes is less than D, the transfer
delay from ingress to egress will be less that D for all packets.
Therefore, the length of the time interval is irrelevant and therefore
superfluous.
The time interval becomes relevant when we look at statistical bounds
(quantiles). I want to remark that packet loss and throughput are as
such statistical bounds, meaning the "probability that a packet is lost
in a time period" and the "average throughput in a time period". If the
time interval is not present, it should get the default value of
infinity.
I quote from the draft:
"The delay and jitter indicate respectively the maximum packet transfer
delay and packet transfer delay variation from ingress to egress,
measured over (any) time period with a length equal to the (indicated)
time interval."
This statement (and all other statements regarding the time interval)
should be changed into
"...time period with a length equal to or larger than the (indicated)
time interval."
The rationale behind this suggestion is the following:
The time interval gives the granularity of the guarantee, meening that
on a shorter time scale, the guarantees are not valid. However, on a
longer time scale the guarantees should hold, otherwise, it would not
make sense to indicate them. An example to illustrate this reasoning:
Imagine a link which has the (strange) behaviour of dropping 1 packet
out of the 50 sent packets every 2 seconds. If we measure the packet
loss on a time scale of 2 second, we will measure an average packet loss
of 2% (1 out of 50). However, if we measure over a time scale of 3
seconds, we will sometimes measure 2.7% (2 lost packets out of 75 sent).
With the current specifications, we can specify a the couple (packet
loss=2%, time interval=2 seconds). However, the maximum packet loss is
measured when 2 out of 51 packets are lost (the first and the last),
yielding 3.9%, so the specified couple should be (packet loss=3.9%, time
interval=2 seconds).
Best regards,
Tim.
-- Timothy Soetens _________________________________________>From the Network Strategy Group @ Alcatel
Tel. +32-3-240.42.11 Fax +32-3-240.48.88
This archive was generated by hypermail 2b29 : Fri Jul 20 2001 - 15:47:53 CEST