RE: draft-ietf-rmt-bb-fec-raptor-object-02 WG Last Call

Vincent & Christophe,

Please see comments below, marked MW:

Regards,

Mark

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of
Vincent Roca
Sent: Wednesday, September 28, 2005 5:32 PM
To: Lorenzo Vicisano; [email protected]
Subject: Re: [Rmt] draft-ietf-rmt-bb-fec-raptor-object-02 WG Last Call

Dear authors and al.

Please find below a few comments for
draft-ietf-rmt-bb-fec-raptor-object-02.txt

   Vincent and Christoph


-------------------

This Raptor scheme requires that block size be less or equal to
Kmax=8192
source symbols. This is relatively small, which is confirmed by  RFC3453
that says, section 2.4:
   Like Tornado codes, the number of source symbols k may be very
   large for LT codes, i.e., on the order of tens to hundreds of
   thousands [...]
Hence the question:

*** comment 1: is this a fundamental limitation or only a practical
    one? For instance, I see that section 5.8 provides an exhaustive
list
    of possible values for K between 4 and 8192, so I can understand
that
    providing a larger table would quickly become tedious. Is that one
    of the reasons for this limitation?

MW: Yes, this is one of the reasons - as well as storing the table in
implementations. Another is that having this limitation means that we
can also limit to 65536 symbols in total (source and repair) and this
keeps the FEC Payload ID small (the FEC Payload ID being an overhead in
every packet) and assists with optimised implementation. If Kmax were
larger then the limit of 65536 symbols in total becomes problematic,
since it places a lower bound on the rates supported.

Tornado, LT and some other LDPC codes require very large numbers of
symbols to achieve efficient performance. In practice, it is desirable
to have the freedom to use larger symbols (for computational reasons),
so the possibility to work efficiently with smaller number of symbols is
a plus point, because it means symbols can be larger (up to the packet
size).

Finally, in practice this limitation means that a very large file may
have to be split into a small number of separate source blocks. Since
these blocks are themselves still very large, the effect of the "coupon
collector problem" is actually very small (the variance in the number of
lost symbols across a few blocks of >8192 symbols will be very small).
-------------------

*** comment 2: is Raptor FEC Enc ID 1 generic or not?

More precisely, is the present FEC scheme sufficiently generic to
be used in many different target environments (e.g. over high speed
Internet connections) or is it specialized for the 3GPP/DVB-H
use cases? I think it has been said that the IETF and 3GPP
specifications
were aligned, but correct me if I'm wrong. The 8192 limitation for
Kmax, which might be appropriate in 3GPP/DVB-H use cases (here also
correct me if I'm wrong), is suboptimal for a general purpose FEC codec
when dealing with large object, since it opens the door to this good
old ``coupon collector problem'' that large block/expandable codes
were supposed to solve. Can you clarify please?

And the corrolary: if ever it is a carefully tailored FEC scheme
for a specific use case, do you plan to submit another Raptor FEC
scheme?

MW: Yes, it is intended to be generic. See above for comments on the
block size limitation. We believe this is the right balance between
supported block size, the overhead of the FEC Payload ID and the need to
provide some limitation for implementation practicality (including
provision of the table of Systematic Indices).

The code is aligned with 3GPP and DVB-H but it is not limited to
3GPP/DVB-H object sizes and supports objects from kilobytes to
gigabytes. Note that it's the object size that is relevant, not the
connection speed - you should not assume that people never want to send
large objects over slow connections ;-)


-------------------

*** comment 3: definitions/symbols/abbreviations

Definitions, symbols and abbreviations only appear in section 5.2,
after they have been heavily used and even re-defined earlier.
A few pathological examples of bad re-definitions are:

Section 4.1.
      - The transfer length of the object, F, in bytes
Then in section 4.2., same page:
      - F the object size, in bytes
Finally, section 5.2.2:
   F    the object size, for object delivery, in bytes

or:

Section 4.1:
      - A symbol alignment parameter (Al)
Then in section 4.2:
      - Al the symbol alignment factor, in bytes
Finally, section 5.2.2:
   Al   denotes a symbol alignment parameter.  Symbol and sub-symbol
        sizes are restricted to be multiples of Al.

I see no reason for delaying these definitions, which makes both
understanding difficult and creates duplicated, non coherent,
pre-definitions. I suggest to move section 5.2 before section 3.

MW: Only the occurance in the "definitions" section is supposed to be a
definition. I'd prefer not to re-arrange sections, but I can fix the
discrepancies in nomenclature.
-------------------

*** comment 4:
What is the motivation for including 13 full pages of numbers in
the main part of the document rather than having them defined in
an Annex, in section 5.8? The same is true for section 5.7.1.

MW: There is no special reason. They could be in an Annex if people
prefer (but not an Appendix ;-)

*** comment 5:
By the way, for my personnal understanding, do the coder and
decoder need to memorize the 8192 - 4 systematic indices values
or not? A carreful reading of the document would certainly help
answering this question, but since the explanations are sometimes
really scarce (in particular in sections 5.5.4 where there are
many ``magic formulas''), I cannot easily answer...

MW: Yes, the encoder and decoder must both know these numbers. They
ensure that the first K symbols of the code are linearly independent and
thus can be systematic symbols.

The specification is intended to rigourously specify the code and you
are right that it does not provide a lot of explanation of why it is the
way it is. I am not sure the formulae are "magic" - they are just the
way they are because that is what makes the code work!

-------------------

*** __Minor__ remark for figure p7

    3         2         1
   10987654321098765432109876543210
   +--------------+-------+-------+
   |      Z       |   N   |   Al  |
   +--------------+-------+-------+

This kind of graphic which shows bit alignment fields requires to
have some space between the bits, so that the separation between the
Al, N fields can be shown as taking place between bits 7 and 8,
rather than exactly at bit 8 (before or after?). In the present case,
common sense enables to easily answer, but adding "+" between each
"-" is preferable anyway.

MW: Ok.
_______________________________________________
Rmt mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/rmt