Re: [Fecframe] Comments aboutdraft-galanos-fecframe-rtp-reedsolomon-00

["Einat Yellin (Lachover)" <Einat at radvision.com>]

Hi Vincent,
Inline.

Thanks again,
Einat

-----Original Message-----
From: Vincent Roca [mailto:vincent.roca at inrialpes.fr] 
Sent: Monday, November 09, 2009 8:10 AM
To: Einat Yellin (Lachover)
Cc: Orly Peck; Sarit Galanos; fecframe at ietf.org
Subject: Re: [Fecframe] Comments
aboutdraft-galanos-fecframe-rtp-reedsolomon-00

Hello Einat,

I had a discussion with Orly today morning and it helped me clarifying 
some aspects.
See my comments below.

> * Section 4 is confusing because it mixes the notions of packets,
> symbols,
> and "elements of the Galois Field". The following is the official 
> terminology:
> [...]
> [Einat] You are correct here. There are some inconsistencies and wrong
> terminology in this section.
> May I suggest:
> In general a Reed Solomon code takes a group of K source symbols and
> generates N - K repair symbols. A receiver needs to receive any K of
the
> N source or repair symbols in order to recover the remaining N-K
> symbols. In this document we suggest a symbol to be a whole RTP
packet.
> As explained in RFC5510, the Reed-Solomon algorithm operates over
> multiple elements each taken from a single source symbol. symbols are
> composed of S "m-bit elements", where m is the Galois Field exponent
> GF(2^m). In the usual case of GF(2^8), elements are bytes, and the
size
> S in terms of elements is of course equal to the symbol size in bytes.
> In terms of implementation simplicity it is also recommended to use
> 8-bits elements.  For more information on Reed Solomon codes, the
reader
> is referred to [Rizzo97].
>   
=> VR: okay.

> Similarly, in section 5, it is said:
>   "A source block for the Reed-Solomon code contains K data blocks."
> There is no notion of "K data blocks" in the official terminology. I 
> think you
> want to say:
>   "A source block [...] contains K source symbols."
> (idem for the following sentences).
>
> [Einat] You are right again. Data block should be replaced with Source
> symbols. To link this to the FEC framework the appropriate phrasing
> should be: 
> A source block for the Reed-Solomon code contains K source symbols.
In
> the scheme presented by this document, each source symbol contains a
> single Application data Unit (as defined in FEC framework), which is
in
> our case RTP packet. Therefore a source block contains exactly K
> consecutive RTP packets.
>   
=> VR: okay

> * Section 5: is the "one symbol per RTP packet" really appropriate?
> If RTP packets have largely different sizes (usual case), the sender
may
> want to use symbols of size the median value (for instance) in order
to 
> avoid
> having too much padding. Of course padding is not sent (i.e. there's
no
> transmission overhead penalty), but encoding/decoding operations will
be
> faster with smaller symbols. There is a price to pay for that,
> essentially
> complexity (it breaks the packet <=> symbol relationship and there are
> more
> symbols), so I'm not sure there's a real benefit. What do you think?
> [Einat] It is true that one could use a median packet size in order to
> make encoding/decoding more efficient. However, since we are dealing
> with packet losses and not symbol losses, breaking a large packet into
> several symbols and unifying small packets to a single symbol will
break
> the equality between packets. For example, if we use 2 repair symbols
> per a block and one of the RTP packets in this block contains 3 source
> symbols then if we lose this single packet we could not restore it
even
> if all other source and repair symbols are received.
>   
=> VR: you're right with your example, but it is a little bit biased. If

we think in terms of
"transmitted bytes" rather than "number of packets", the comparison is 
not fair. If the
symbol size is half the size of the largest source packet, then I can 
send two repair packets
for almost the same transmission overhead as what you are proposing in 
the I-D (I'm not
considering here the various protocol header overheads of course).

There's an additional incentive for using median symbols sizes that I 
forgot to mention:
ff symbols are of size the largest source packet size, it means that the

repair symbols are also
of that size, which has an impact on the transmission overhead. With 
median size symbols,
this would not be the case, and it makes a big difference if there are a

significant number
of repair symbols.

I note there's a misunderstanding in your answer: I don't suggest to 
"unify small packets to
a single symbol". I still padd source packets that are smaller than the 
symbol size.
[Einat] Rethinking this - 
if no unifying of small blocks and with the correct protocol to back it
up this could be a good idea.
Anyway, it will be left for the transmitter to decide about the symbol
size.

> * Section 5: It is said:
>       "B.  The sequence number of the source packets must be
consecutive
>            in a source block."
> IMHO the "MUST be consecutive" is too strong. It may happen that some
> source RTP packets be lost in some networks prior to reaching the
FECFRAME
> sender (i.e. if FECFRAME is used in an intermediate box rather than
> end-to-end).
> And what about (crazy) situations where RTP packets are lost (e.g. by
a
> FIFO overflow) before reaching FECFRAME?
> IMHO it's worth designing a solutions that is robust against this kind
> of situation, no matter the cause.
>   
> [Einat] The 'MUST be consecutive' restriction is there in order to
> simplify the signaling of which RTP packets are protected. If not
> consecutive then we could not use the SN base + num packets fields in
> the FEC header and we'd have to use SN base + bitmask, or worse -
> specifically list the sequence numbers of protected packets. I think
> that cases where some source packets are lost before FEC encoder are
> unique and can be handled by creating smaller FEC blocks (in between
the
> gaps).
>   
=> VR: I understand the simplification purpose. With end2end FECFRAME 
(i.e. located
on the same host as the source application, it's a reasonable 
assumption. In other situations,
it makes the proposal unusable, unless you have a TCP (or similar) 
connection to guaranty
there's no loss. So I'm in favor of some solution. Which solution to 
choose is a separate
discussion (see below).

> * Section 6.2.1:
> One stupid question: if there are several repair flows and if they are
> all sent to the same multicast group/port, how does a receiver tell
the
> difference between the packets belonging to the different repair
flows?
> I think it's by means of the PT. This is not said in any case.
> (I understand that sending everything to the same destination is not
> necessarily a good idea, but that's a different topic)
> [Einat] The payload type, which is signaled in SDP is used for that.
We
> could add a specific sentecnce describing this.
>   
=> VR: yes, please.

> * Section 6.2.2, Fig. 3: the choice of the various field sizes has
> implications that are not explicitly mentioned.
> Num Packets (badly named, I'd prefer K): 16 bits => too much in case
>         of RS over GF(2^8), more appropriate to GF(2^16).
> [Einat] We did not limit GF to 8. It is only a recommendation.
> i and N-K: 8 bits => okay with GF(2^8), a limit in case of GF(2^16),
> but that's perhaps not the target.
>
>
> * Whole document: More generally, I have the feeling nowhere it is
> explained 
> whether or not there are restrictions on the GF size. Which possible
> values
> for the m parameter.
> [Einat] I agree that such restrictions and consideration should be
> further elaborated.
> In general we focused more on the concept of how to apply Reed-Solomon
> on RTP packets than on the Reed-Solomon operation.
>   
=> VR: okay, but an I-D that specifies actual header formats must be 
clear on such
assumptions.

> * Section 6.2.2: I have the feeling that carrying N-K in the FEC
header
> is
> perhaps not required. On the opposite, there MUST be a way to inform
the
> receiver of the GF(2^m) used (e.g. it can be done out-of-band in SDP 
> description).
> So if I know m, I don't need to know the actual N-K as I know K and 
> max_K=(2^m)-1.
> [Einat] I'm not sure that I follow here. Our intention was to use a
> dynamic N-K value. 
> e.g. for one block it could be 2 repair symbols and for the next block
> it could be 4 repair symbols.
> This variable number must be signaled for each block.
>   
=> VR: I misunderstood your point and the discussion I had with Orly 
today clarified
a lot. So let's summarize:
- for a pure RS encoding/decoding aspect, knowing n-k is not needed, m 
is sufficient
  (it was the main motivation for my comment);
- from a practical point of view, sending N-K can be helpful in some 
situations (but
  not all) to help the receiver take decisions on whether or not 
decoding will be feasible.
  It makes sense. However (1) it requires the receiver actually got at 
least one repair
  packet otherwise he won't know if it's worth waiting for hypotetical 
repair packets,
[Einat] In the scheme presented the receiver must receive at least one
repair packet in order to be able to repair, not only to be able to know
if he can repair.
So the receiver is required to wait and if wait time elapses (how long?
This is another issue) and there is no repair packet received for this
packet than it can assume that this packet cannot be restored (either it
was not originally protected or its repair packets were lost).
  and (2) it's not absolutely required, we can also use the realtime 
deadline of the data
  transported.
All in all, carrying N-K can help, but I don't think it's the ultimate 
solution, we also
need to consider the realtime deadline of data being transported for 
robustness.

> * Section 6.2.2: SN Base is used to identify the source block (since
all
> the FEC repair packets associated to the same source block have the
same
> SN Base
> value). Since SN Base is a field that regularly wraps around, there is
a
> (theoretical) risk that two consecutive source blocks use the same SN
> Base
> value. This might happen in tricky configurations with large source
> blocks
> (GF(2^16)) and in case the RTP sequence numbers are not consecutive
(see
> above
> discussion about possible erasures *before* FECFRAME).
> Additionally, and more seriously, this vulnerability may also be 
> exploited as
> a possible DoS (Denial of Service) to make a FECFRAME sender crash.
> So it's worth mentioning this topic in the doc, both in the core spec 
> and in the
> Security section.
> If we clarify that this document is not for situations experiencing
high
> erasure
> rates before FECFRAME (which can be checked on-the-fly), it should be 
> okay IMHO.
> [Einat]  Good point. Thanks.
>
> * Section 6.2.2: Back to the idea of removing the constraint on
> consecutive
> source packets, I'm wondering if the following FEC header format
> wouldn't be
> okay:
>
>  0                   1                   2                   3
>  0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
> +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> |          Base SN              |           Max SN              |
> +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> |    Source Block Length (k)    |   Enc. Symbol ID (16 bits)    |
> +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-++-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
>
> Thanks to this FEC header:
> - We know that all SN values in {Base SN..Max SN} are included in this
>   source block.
> - We know their actual number (k) even if some of RTP SN are missing.
> - We know the ESI of the FEC repair packet (in {K, N-1}), no matter
the
>   GF m parameter since it's 16-bit long.
> - And we know in which order incoming source RTP packets should be
>   stored in the block, even if there are gaps in the SN.
> At a receiver, the actual SN value in the RTP header of an erased
> source packet is anyway recovered along with the rest of the
> header/payload
> so there's no problem either.
> [Einat] This is interesting. Do you mean that all RTP packets in the
> range [Base SN...Max SN] will be included in this block?
> If not then the receiver will know for sure if he can recover a lost
> packet only after making the decoding operation on the relevant block.
> This could lead to unnecessary decoding calls.
> See my comment above regarding the need for signaling N-K for each FEC
> block.
>   
=> VR: my above proposal is not sufficient! It won't help identifying 
the position of source
packets inside the source block.

However, I think there's a solution to the case where there's a limited 
number of erasures
prior to FECFRAME. Instead of carrying a bit mask that identifies which 
source packet
of a SN range are actually considered or not, I suggest to replace 
erased source packets
by zero'ed symbols.
These symbols will be reconstructed by the receiving side FECFRAME, and 
a quick
check will enable him to get rid of them immediately (there's no RTP 
header, it's all zero).

To answer your question: yes, all the packets in [Base SN; Max SN] are 
in the block,
except that some of them are "fake" packets (zero'ed) which means they 
are easily
recognized.
[Einat] If I understand what you suggest then increasing the source
block size as you proposed will increase encoding/decoding complexity.
So there's a tradeoff to consider here.

> * Section 7.1.1: it is not clear to me if max_N is equal to 2^m or if
> it is only a "codec limitation". I think it's related to the GF, but
> it's
> worth mentioning. And the "symbol-size" should also clarified as
already
> mentioned.
> [Einat] Symbol size should be element size as mentioned above. 
> I also agree that considerations for determining MAX_N should be
listed.
>
>
> * Section 8.3.1: This section does not explain how to manage the case
> where
> there are several repair flows. It's a problem.
> [Einat] What do you see as a problem here? The different repair flows
> are distinguished by their payload type. 
> It is determined in the SDP which payload type is used by which repair
> flow to protect which RTP flows.
>   
=> VR: said like that it's clear. But it just needs to be said (same 
comment as above).

> * Section 11.
> You cannot get rid of the security aspects in this way. Most
> specifications
> add additional security risks (like the DoS opportunities mentioned
> above).
> [Einat] Correct again. Would it be OK to say that we neglected some
> edges in order to make it on time for the meeting...?
> This will be fixed if/when the concept presented here is agreed.
>   
=> VR: okay, I can also help.

Regards,

   Vincent