Before I say more, I think I would like to prefix this with my own
personal view:
Many applications using TFRC will not normally try to send at the
maximum permitted rate (i.e. are data/application limited). I expect
these to be typical of apps expected to use DCCP CCID-3. That is, TFRC
for DCCP could be thought of as providing a congestion-responsive
protocol that responds to congestion in a way that is not (much) worse
than that of TCP - to me this does not imply a need for equality in
terms of throughput with TCP (i.e. I think a good use of the algorithm
is to prevent a TFRC application from sending more, or at least not much
more than a TCP flow would have sent).
Now to the response:
TFRC has been put forward for a wide range of applications - from
extremely low rate (e.g. vocoded VoIP) to high rate (>> Gbps), with LAN
delays to wide-area/wireless environments with long RTTs. I suggest, at
least in the immediate future, most common DCCP applications will
operate at low (kbps) to medium rates (few Mbps). My suggestion is
therefore aim for a good stack for low-medium rates, were course clock
granuality may be OK, and which avoids some of the pitfalls identified
by Gerrit's email. Does anyone have a different vision of the near future?
I read RFC 5348, s4.6 as saying you can send extra data after being data
limited/idle, but shouldn't send long bursts - my thinking was that
this was intended to address an issue for long delay paths. Short RTT
paths anyway allow the sender to rapidly grow the rate, and are not
usually so constrained by TFRC (at least at typical application rates).
Gorry
Ian McDonald wrote:
OK. I'll add a few comments even though I'm a little bit rusty...
I was previously an advocate of low resolution timers and then use
bursts as needed to achieve the average rate as specified in RFC3448.
The reasoning for this was very much as you discuss in point 1 - that
you achieve less than the desired rate with high resolution timers as
you will never get exactly to transmit at the time you require (unless
you have a "hard" realtime system with desired accuracy) - so any
delay will slow down your transmit rate, and that high resolution
timers may not be available on all architectures. I also had a third
reason - overhead - if you're interrupting other tasks and having to
do a context switch many, many times a second surely that isn't so
good?
However RFC 5348 changes this as this clause is added to 4.6:
To limit burstiness, a TFRC implementation MUST prevent bursts of
arbitrary size. This limit MUST be less than or equal to one round-
trip time's worth of packets. A TFRC implementation MAY limit bursts
to less than a round-trip time's worth of packets
and this is further explained in section 8.3 and the downside - that
you can't send big bursts so you can't get the full calculated rate.
The RFC uses an example of 1 msec scheduling and 0.1 msec RTT. However
what would be worse is devices on a LAN with 10 msec timer - e.g. two
embedded devices at home - I haven't done the maths but I think the
rate achievable would be quite low.
One thing that I think we do need to be careful about though is
assuming that we should be trying to get very high speed transfer -
DCCP is not what we would layer a file serving protocol on top of....
(some have argued you shouldn't even use TCP for this on a LAN...)
Thinking laterally there is another possible solution - something I
used way back in the 80s for another project - build your own
scheduler! We could set a high resolution timer to tick every 0.1 msec
and then use the coarse grained algorithm at that point....
This is a hack to some degree and I can imagine David Miller
suggesting that it is more a protocol issue... The other thing is that
if we did this we would have to only do it when we actually need and
use higher granularity at other times or else the Powertop people may
not be so happy.
Anyway - something to think about. I've also added the IETF list as
well in case people there have the answers.
Regards
Ian
On Sat, Nov 15, 2008 at 11:50 PM, Gerrit Renker <gerrit@xxxxxxxxxxxxxx> wrote:
I would appreciate some advice and insights regarding the use of
high-resolution timers within a transport protocol, specifically
DCCP with CCID-3 (RFC 5348).
Currently the implementation is in a limbo of high-resolution and
low-resolution code. It is not good, neither here nor there, so
I would like to work on making the interface consistent.
After thinking this through I encountered a number of points
which made me question whether high-resolution timers will lead to
better performance and a cleaner interface.
I'd appreciate comments and input on this, the points are below.
1. Handling unavoidable waiting times
-------------------------------------
One can not expect that, if the scheduling clock says 'send in x
microseconds', a packet will indeed leave after x microseconds; due to
waiting times. An example is in net/dccp/timer.c, when the socket is
currently locked - we wait for a "small" amount of time:
bh_lock_sock(sk);
if (sock_owned_by_user(sk))
sk_reset_timer(sk, &dp->dccps_xmit_timer, jiffies + 1);
else
dccp_write_xmit(sk, 0);
bh_unlock_sock(sk);
2. Dependency on high-resolution timers
---------------------------------------
Committing the CCID-3/CCID-4 implementations to using high-resolution
timers means that the modules can not be built/loaded when the kernel
does not offer sufficient resolution.
This has recently made it hard for someone using CCID-3 to find out
why DCCP would not run, where the cause was that high-resolution timers
were not enabled in the kernel.
3. Noise in the output
----------------------
When tracking the speed of a car every 10 seconds, there is a lot of variation
in the values, due to stopping at traffic lights, accelerating etc. But when
considering a larger timescale, one can say that the average speed from city
A to city B was xx mph, since the whole journey took 2.5 hours.
The same can currently be observed with X_recv - there is one commit which
tries to make X_recv as fine-grained as possible, it is labelled "dccp ccid-3:
Update the computation of X_recv",
http://eden-feed.erg.abdn.ac.uk/cgi-bin/gitweb.cgi?p=dccp_exp.git;a=commitdiff;h=2d0b687025494e5d8918ffcc7029d793390835cc
The result is that X_recv now shows much wider variation, on a small timescale
there is a lot happening. It can best be seen by plotting the X_recv using
dccp_probe. Without this commit the graphs are much 'quieter' and just show
the long-term average.
In TCP Westwood for instance a low-pass filter is used to filter out the
high-frequency changes in the measurements of the Ack Rate:
"TCP Westwood: Bandwidth Estimation for Enhanced Transport over Wireless Links"
Mobicom 2001
http://www.cs.ucla.edu/NRL/hpi/tcpw/tcpw_papers/2001-mobicom-0.pdf
I'd appreciate opinions on this, as I think
With regard to CCID-3, it also seems to be be better to revert the above
commit and just use long-term averages.
4. Not sure using high-resolution is the answer
-----------------------------------------------
While a fine-grained timer resolution may be desirable, it is not
necessarily a must. The implementation of rate-based pacing in TCP
(http://www.isi.edu/~johnh/PAPERS/Visweswaraiah97b.html) for instance
also used low(er) resolution timers and it worked.
The RFC for CCID-3 (http://www.rfc-archive.org/getrfc.php?rfc=5348) also
does not high-resolution; it supports coarse-grained timestamps (section
6.3 and RFC 4342) and discusses implementation issues when using a
lower resolution (section 8.3).
The counter-argument could be that CCID-3 is a transport protocol with a
built-in Token Bucket Filter so that similar considerations apply as for
the Qdisc API (net/sched/sch_api.c).
Summing up, I have doubts that basing CCID-3 will bring advantages and
would much rather go the other way and (consistently) use lower resolution.
Thoughts?
--
Web: http://wand.net.nz/~iam4/, http://www.jandi.co.nz
Blog: http://iansblog.jandi.co.nz
