Conditions when t_ipi, t_nom, t_delta need to be recalculated
"When is it necessary to update t_ipi, t_delta, t_nom" ?
=============================Following Ian's email from yesterday, I spent time working through the RFCs and compiled information about the dependencies among the different parameters. This email is a write-up of what I found and serves as basis for subsequent patches.
The parameters are:
s: sender packet size in bytes, known to sender
X: current sending rate (bytes/second), known to sender
p : loss event rate, undefined until feedback has been received t_RTT : round-trip time estimate, undefined until feedback has been received t_RTO : retransmission timeout, undefined until feedback has been received X_calc: = f(s, p, t_RTT) - also undefined until feedback has been received X_inst: instantaneous xmit rate, f(X) - undefined until feedback has been received
t_nfb: nofeedback time-out, known to sender
t_tld: time-last-doubled, -1 until first packet has been sent
1) Definition of t_ipi
----------------------
In [RFC 3448, 4.6], t_ipi is defined as t_ipi = s/X_inst
If the feature `Preventing Oscillations' [RFC 3448, 4.5] is not implemented, then X_inst = X. Otherwise, X_inst depends on the feedback sample for t_RTT and is undefined until the first feedback has arrived. The assumption here is that the optional `Preventing Oscillations' feature is not implemented, as it requires square-root computation which is complicated without floating-point arithmetic. Then,
t_ipi = s/X
and X can be (1) the initial value or (2) the recalculated value after feedback has been received. Hence t_ipi needs to be updated whenever
(i) s changes
(ii) X changes
To simplify, we assume that s' is fixed and therefore t_ipi only needs to be updated whenever X changes. However, in the actual patch (which tracks the mean value of the packet size in s'), a note will be added as a possibility for further optimisation (it is also not implemented since the standard is silent about whether the t_ipi should be updated whenever the mean value `s' changes).
2) What depends on t_ipi
------------------------
The following values are derived from and thus depend on t_ipi [RFC 3448, 4.6]:
* nominal sending time: t_nom = t_(i+1) + t_ipi
* scheduling delta : t_delta = min(t_ipi/2, t_gran/2)Another dependency is the nofeedback timer: after initial feedback has been received, it is set to
t_nfb = max(4*t_RTT, 2 * s/X)
= max(4*t_RTT, 2 * t_ipi) Thus these three values, in turn, need to be updated when t_ipi changes.
3) Initial values
-----------------
Initial settings are described in [RFC 3448, 4.2]:
X/s = 1hz
t_RTT, t_RTO are undefined
t_tld = -1
Thus we have the following initial values:
t_ipi = s/X = 1 second
t_nom = t_now + t_ipi
t_delta = t_gran/2 (since OS schedules finer than 1 second)
==> A patch has already been sent for this case.
4) Sending until the first feedback packet arrives
--------------------------------------------------
No need to update: X, t_ipi, t_delta, t_nfb stay at their values, t_nom is incremented in the manner
t_nom += t_ipi [where t_ipi remains constant]
=> This is already implemented by ccid3_hc_tx_send_packet
5) When the first feedback packet arrives
-----------------------------------------
The first feedback packet sent by a receiver has p = X_recv = 0 [RFC 3448, 6.3]. Plugging these values into step (4) of [RFC 3448, 4.3] and considering t_tld = -1, gives
X = max(2 * min(X, X_recv), s/R)
= max(2 * min(X, 0), s/R) = s/RWithout modification, the first sending rate remains at 1 packet per round-trip time. [RFC 4342, sec. 5] increases this to 2 * s/R, up to the maximum of 4 packets per RTT and 4380 per RTT (this assumes that MSS-sized packets are sent).
==> A patch is required to consider this special case.
6) When the second ... n-th feedback arrives
--------------------------------------------
X, t_ipi, t_delta, t_nfb, t_nom need to be updated.
==> This is already handled by ccid3_hc_tx_packet_recv.
7) When the nofeedback timer expires and no feedback has been received so far (step 4.4)
----------------------------------------------------------------------------------------
X = max(X/2, s/t_mbi)
--> t_ipi, t_nom, t_delta must be recalculated
t_nfb = max(4*t_RTT, 2*t_ipi)
==> This needs updating, there is a FIXME in ccid3_hc_tx_no_feedback_timer at this place.
8) When the nofeedback timer expires and feedback has been received previously (step 4.4)
-----------------------------------------------------------------------------------------
* t_RTT, t_RTO, and p do not change with regard to the last-received feedback
* thus X_calc is as before
* X_recv gets changed
* hence X gets changed as a result of step (4) of [RFC 3448, 4.3]
--> t_ipi, t_nom, t_delta, and t_nfb need to be updated
==> This also needs a patch, since X is updated, but t_ipi, t_nom, and t_delta are not
| * nominal sending time: t_nom = t_(i+1) + t_ipi
=> should read t_nom = t_i + t_ipi
Quoted 8 lines · Gerrit Renker
"When is it necessary to update
t_ipi,t_delta,t_nom" ?
=============================Following Ian's email from yesterday, I spent time working through the RFCs and compiled information about the dependencies among the different parameters. This email is a write-up of what I found and serves as basis for subsequent patches.
This is really excellent work Gerrit.
I think we should put this on the ODSL Wiki or in a text file under Documentation.
Ian
Signature · 7 lines
Ian McDonald Web: http://wand.net.nz/~iam4 Blog: http://imcdnzl.blogspot.com WAND Network Research Group Department of Computer Science University of Waikato New Zealand