[ippm] comments on draft-ietf-ippm-framework-compagg-04

["Loki Jorgenson" <[e-mail]>]

Al - some comments on draft-ietf-ippm-framework-compagg-04

Alan Clark's comments from last week are very relevant - in particular
w.r.t. appropriate statement of assumptions related to implicit models
and methodologies.

In addition (and somewhat redundantly since I had already written the
notes):

General:
  o of course, choice of PDV and IPDV as relevant implementations seems
valid - however, there are implicit assumptions that should be made more
explicit.  E.g. the values for reference time embody particular choices
related to methodology and model:
          o PDV assumes minimum delay knowable and difference optimally
significant (transmission distribution is not relevant)
          o IPDV assumes, or is typical within the context of, periodic
transmission (or some context wherein packet separation is significant)
and that the regularity of the period represents an optimal reference
(and propagation time is not specifically relevant)

Of course these descriptions are intended to be considerably more
generic but in doing so they look past the obvious implementations
(which is part of the intention of this draft, yes?).  Some of the
descriptions of RFC 3933 include references to other timing
distributions (e.g. Poisson).  The choices of minimum and delta are
specific and motivated (by application requirements).

  o references to "optimizing de-jitter buffer" are a relevant example
of application design references that motivate the choice of jitter
definition - is that related to the (anticipated) mandate of PMOL/APM?

  o There should be more discussion of the typical implementation
requirements - there is only passing mention of:
           o optional statistical approaches
                   - holding and post-processing all data in a
measurement interval
                   - not holding (much) data and maintaining running
measures
                   - holding a sliding window of data and maintaining
sliding measures
           o memory requirements (based on each case above)
           o implications for reference values (based on each case
above)
                   - e.g. D(min) as running value for whole set or
absolute for windowed set
           o issues and complication possible in various instances of
each case
                   - e.g. quantiles are not available as running measure
         ** many of these issues may be adequately addressed by simply
emphasizing the MAPDV2/G.1020 document.  There is some mention of these
issues in Appendix 13.

  o "quick variation" seems an unfortunate turn of phrase
           o however, it has been defined as being on the order of
inter-packet separation so that seems adequate
           o an alternative might be "short timescale variation"

   o in the definition of IPDV, why is there no consideration of the
practical context wherein packet arrival times are known but not
transmission times for the periodic case.  Is that to suggest that there
is no meaningful implementation of IPDV without transmission times?
          
   o in Sec. 8.1 referring to Measurement Stream Characteristics, there
should be some description of the impact of sampling methodology for the
cases wherein not all the packet D(i) are sampled.  This should include
some suggestion for methodology references.
            o for example: 
The implication of short-range dependency on delay variation measurement
Qiong Li; Mills, D.L.
Network Computing and Applications, 2003
Volume , Issue , 16-18 April 2003 Page(s): 374 - 380
	o for example
Fisher Information of Sampled Packets: an Application to Flow Size
Estimation
Bruno Ribeiro, University of Masachusetts at Amherst; Don Towsley,
University of Masachusetts at Amherst; Tao Ye, Sprint Advanced
Technology Laboratories; Jean Bolot, Sprint Advanced Technology
Laboratories

   o in Sec. 8.3 Test Duration, some references to stability of networks
would be useful.  Although changes are possible on almost any time
scale, there are some reasonable rules of thumb to be considered.
Typical networks are stable on the order of minutes.  However there are
also long timescale variations associated with daily cycles.
           o for example
On the Constancy of Internet Path Properties. 
Yin Zhang, Nick Duffield, Vern Paxson, and Scott Shenker -- AT&T Labs
Research, Florham Park NJ; AT&T Center for Internet Research at ICSI,
Berkeley CA.

   o in Sec. 8.5, there is reference to distinguishing long delays from
loss - as mentioned elsewhere, a reasonable threshold should be applied.
Naturally, discards should be accounted for distinct from loss.

  o in Appendix 12, also include various forms of equipment and
configuration dysfunction as a separate cause.  For example, consider
re-transmission effects in WiFi or slow-path processing in a router that
has been sub-optimally employed for edge filtering.

  o in Appendix 13, consider including code fragments for the described
method.  This may be a special case of temporal composition (associating
sets of data segmented intervals from a single testing period).

-- Loki 

_______________________________________________
ippm mailing list
[e-mail] 
https://www1.ietf.org/mailman/listinfo/ippm