[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[ippm] Comments on draft-ietf-ippm-framework-compagg-04.txt



Al
 
The draft has some good definitions of metrics - however to make it a little more generic and hence capable of wider application it may be useful to make some minor changes.
 
(i) Composite Metric definition - 1
There are many cases in the industry where the term Composite Metric is used.  One common case is where a metric is constructed from several dissimilar metrics - for example a simple case is the combined loss/discard rate in RFC3611, which is not the sum of two packet loss rates but the sum of a packet loss rate with a jitter buffer packet discard rate.  There are many more complex cases in common usage.  The text at the present seems to be more oriented towards composition of identical metrics - either in space or time.
 
Recommendation: add text to state that composite metrics may be constructed from either identically defined metrics or from dissimilar metrics.
 
(ii) An "Index"
There are common cases when an artifical index is created - often from a complex function of some metrics.  The value range of the index typically has no relationship to any observable parameter - and is selected for convenience, the function use to calculate the index is then determined.  Obvious telecom examples are MOS scores and R factors. 
 
Recommendation:  Add
3.n Index
An Index is a composite metric for which the output value range has been selected for convenience or clarity, and the behavior of which is selected to support ease of understanding.  The composition function for an index is often developed after the index range and index behavior have been determined.  Examples include MOS scores and R factors.
 
 
(iii) Models and metrics
Many metrics are either implicitly or explicitly the result of assuming a model of the process being measured, and then estimating the parameters of the model.  Simple examples include "average packet loss rate", which seems to assume independent loss probability, RFC3550's PPDV, which seems to assume alternately early (or on-time) and late packets, and consecutive loss metrics.  IMHO it is better to explicitly state what model is being used, as that brings out the assumptions being made and allows their validity to be questioned.  For example, one of the most common means of reporting performance is to measure the average of some value during a time interval - in many cases this is done simply because it is a common thing to do and the process of analyzing the behavior of the parameter being measured is skipped.   
 
Models can also be very useful in understanding system behavior and identifying particularly useful metrics.  I've used the term "intermediate model" to mean a model that is (a) representative of some key aspect of performance for an/the application that is using the service/ system being measured and (b) which is capable of accepting as input data extracted from the service/ system being measured.  In some senses an intermediate model is a composition function - however the function has been designed to be "tuned' for some class of applications.  One example of this is the 4-state Markov Model used in RFC3611, which captures information on sparse bursts of lost or discarded packets, another is the MAPDV2 jitter metric, which is based on a simplified representation of jitter buffer behavior.
 
Recommendation:
 
Add to Section 5. text to indicate that the model (statistical) of the process being measured and the validity of this assumption should be explicitly stated
 
Add;
 
Section 4.n Composition Functions based on Intermediate Models
An Intermediate Model is a model that is representative of some key aspect of performance for a class of applications and which is capable of accepting
 
_______________________________________________
ippm mailing list
ippm at ietf.org 
https://www1.ietf.org/mailman/listinfo/ippm