The HTB qdisc has a compile time option, HTB_HYSTERESIS,
that trades accuracy of traffic classification for CPU
time. These patches change hysteresis to be a runtime
option under the control of "tc".
The effects of HYSTERESIS on HTB's accuracy are significant
(see chapter 7, section 7.3.1, pp 69-70 in Jesper Brouer's
thesis: http://www.adsl-optimizer.dk/thesis/ ), whereas
HTB's CPU usage on modern machines using broadband links
is minimal. Currently HYSTERESIS is on by default, and
requires a kernel re-compile to change. Altering it to
be a runtime option will make life easier for the bulk of
its users.
Further documentation on the patch and its usage can be
found here:
http://www.stuart.id.au/russell/files/tc/tc-atm
This is a combined effort of Jesper Brouer and Russell
Stuart, to get these patches into the upstream kernel.
Let the discussion start about what we need to change to
get this upstream?
We see this as a feature enhancement, as such hope that
it can be queued in davem's net-2.6.18.git tree.
--
Regards
Russell Stuart and Jesper Brouer
Russell Stuart wrote:
The HTB qdisc has a compile time option, HTB_HYSTERESIS,
that trades accuracy of traffic classification for CPU
time. These patches change hysteresis to be a runtime
option under the control of "tc".
The effects of HYSTERESIS on HTB's accuracy are significant
(see chapter 7, section 7.3.1, pp 69-70 in Jesper Brouer's
thesis: http://www.adsl-optimizer.dk/thesis/ ), whereas
HTB's CPU usage on modern machines using broadband links
is minimal. Currently HYSTERESIS is on by default, and
requires a kernel re-compile to change. Altering it to
be a runtime option will make life easier for the bulk of
its users.
At time of HTB implementation I needed to reach 100MBit speed on
relatively slow box. The hysteresis was a way. On other side I used
hand-made TSC based measure tool to compute exact (15%) performance
gain. Today I'd measure it using oprofile.
When rethinking it again I'd suggest to re-measure real performance
impact for both flat and deep class hierarchy and consider switching the
hysteresis off by default (or even to remove the code if the gain is
negligible). If it is the case then it is the cleanest solution IMHO.
On other side I see no problem with attached patches. Have you tested
patched kernel with old "tc" tool ?
thanks for your effort,
Martin
On Thu, 2006-06-15 at 11:49 +0200, Martin Devera wrote:
At time of HTB implementation I needed to reach 100MBit speed on
relatively slow box. The hysteresis was a way. On other side I used
hand-made TSC based measure tool to compute exact (15%) performance
gain. Today I'd measure it using oprofile.
When rethinking it again I'd suggest to re-measure real performance
impact for both flat and deep class hierarchy and consider switching the
hysteresis off by default (or even to remove the code if the gain is
negligible). If it is the case then it is the cleanest solution IMHO.
I attended LCA 2006 this year. There was a presentation by
a group in New Zealand using Debian running on a embedded
box to bring the Internet to rural communities. Some of
these communities didn't have power or telephone, so the
setup ran over 802.11 over distances of up to 23Km using
solar cells for power. I don't recall exactly, but I think
the embedded box was using a 486 equivalent. I think they
had around 40 of these things up and going.
The point of the story is there are people out there who
use Linux on small processors, and often do imaginative
things with them. We would be doing them a disservice by
ripping out the code.
On other side I see no problem with attached patches. Have you tested
patched kernel with old "tc" tool ?
Yes.