[e2e] Observations on RFC 3448 (TFRC)

[Dado Colussi]

Hi folks,

I went through the RFC 3448 (TFRC) in quite detail recently and made
some observations I thought I might share here.


Section 4.4 Expiration of nofeedback timer:

         <quote>
         If (X_calc > 2*X_recv)
         </quote>

         X_calc can be computed only when p > 0. The given code
         fragment assumes a valid X_calc.


Section 4.6 Scheduling of Packet Transmissions: on 4th line:

         <quote>
         course-grain
         </quote>

         Should be coarse-grain?


Section 5.5 History Discounting:

         The elements from 1 to n of DF_i array are initialized to 1.
         DF_0 is left uninitialized. It is, however, referred to in
         the last code fragment when a new loss event occurs:

         <quote>
         for (i = n-1 to 0 step -1) {
                 DF_(i+1) = DF * DF_i
         }
         </quote>

         DF_1 is set to DF_0 which is effectively a random value.
         DF_0 is initialized to 0 only right after that. Thus, the
         random value traverses through the DF_i array until discarded
         after n loss events.

Section 6.2 Expiration of feedback timer, bullet 2:

         <quote>
         Calculate the measured receive rate, X_recv, based on the
         packets received within the previous R_m seconds.
         </quote>

         Feedback is sent approximately once per round-trip time, if
         data packet have been received since last feedback. If the
         data rate is less than one packet per round-trip time,
         feedback is sent only at "rounds" where data messages have
         been received. However, X_calc is calculated over a R_m
         seconds.

         Now, if the data rate is less than one packet per round-trip
         time, X_calc is calculated using one packet over a time frame
         of R_m seconds. For a source sending a packet every fourth
         round, the receiver calculates X_recv four times larger than
         it really is.

         How about calculating X_recv over the time between two
         consecutive feedbacks?


Section 6.3 Receiver initialization:

         <quote>
         When the first packet is received:
         ...
         o Set the feedback timer to expire after R_i seconds.
         </quote>

         In section 4.2 Sender initialization the sender's estimate
         of round-trip time time is explicitly left undefined until
         measured. Thus, the first packet contains an undefined
         value as round-trip time estimate and the feedback timer
         gets initialized to an undefined time.

         I solved this by sending feedback for each packet until
         the sender provided a valid estimate.


Section 6.3.1 Initializing the Loss History after the First Loss Event

         <quote>
         ... the TFRC receiver calculates the loss interval that would
         be required to produce the data rate X_recv, and uses this
         synthetic loss interval to seed the loss history mechanism.
         </quote>

         The document is not being explicit how the seeding is done.
         The NS-2 implementation appears to seed the first element of
         the loss history array. I first interpreted this so that all
         n elements are to be seeded.

         Another issue that crossed my mind is related to ECN. What
         if the very first packet contans ECN bit?


Best regards,

Dado