[Cerowrt-devel] cerowrt 3.10.6-1 AQM and ADSL

[Sebastian Moeller]

Hi All,

please find attached my last set of modifications for the AQM scripts. As far as I can tell now the link layer adaptation mechanisms are mutually exclusive, and the values from the GUI are actually passed to the simple.qos script (I did not test simplest.qos, but assume it also works). I would appreciate other ADSL users to test these extensively so we hopefully can put this issue to rest soon (I would prefer to have this settled before I leave ADSL and head for VDSL2 pastures...)

 My testing using RRUL to a host nearer than either NY or CA shows the following average ping times:
AQM		Linklayer Mechanism		linklayer		estimated avgRTT
none		none					none		300ms
on			none					ethernet		80ms
on			htb_private				ethernet		80ms
on			htb_private				adsl			80ms
on			tc_stab					ethernet		80ms
on			tc_stab					adsl			40ms

(All tests performed with shaping uplink to 95% of link rate and downlink to 90%, using the empirically confirmed per packet overhead of 40bytes)

So to my layman's eyes it looks like my assessment (from looking at tc and kernel sources no less) the the htb_private mechanism is currently non-functional seems correct. The tc stab mechanism however seems to be able to actually improve the interactivity as measured by the ping probe noticeably. It also reduces the avg transfer rates for up and downlink, but that is to be expected as proper link layer adjustment increases the apparent size (especially of small packets) so the ping probes eat more of the available bandwidth. That said, even with tc_stab the RTT graph is not as smooth and flat as I would like, it is just much better than the rest :)
	Executive summary: ADSL1 and ADSL2+ users should use the tc_stab methods, and kernel experts should look at why http_private does not work (and ideally someone would hook up that HTB private configuration method in tc to use the same implementation as the stab method, since a) stab is generic, i.e should work with other discs than HTB and b) the kernel fudges the packet sizes already so the additional bit of lying stab performs should not matter. )


Best Regards
	Sebastian
-------------- next part --------------
A non-text attachment was scrubbed...
Name: usr.zip
Type: application/zip
Size: 13573 bytes
Desc: not available
URL: <https://lists.bufferbloat.net/pipermail/cerowrt-devel/attachments/20130815/4b06551f/attachment-0002.zip>
-------------- next part --------------