On Wed, 2012-03-28 at 16:29 -0700, Dave Taht wrote:
> On Wed, Mar 28, 2012 at 3:36 PM, David Woodhouse <[email protected]>wrote:
> 
> > Resolves https://dev.openwrt.org/ticket/11192
> >
> > You can now configure devices to be used as a TEQL slave as follows:
> > # uci set network.ppp0.teql=teql0
> > # uci set network.ppp1.teql=teql0
> > # uci commit
> >
> 
> I am pretty intensely curious as to how the debloating stuff is going to
> work with bonded adsl lines.

If you apply your SFQ qdisc to the 'teql0' interface, it should be
fairly much identical. The upstream packets are sent over both
interfaces, in an alternating fashion — but apart from the occasional
reordering of packets it's not really that different to having a single
line with double the bandwidth.

Speaking of bloat, I noticed that there's a *huge* "hidden" queue of
outbound packets on each PPPoATM interface. It'll buffer up to the
socket send buffer (/proc/sys/net/core/wmem_default) between the generic
PPP core and the ATM device. That's 160KiB on my system. Per interface.
Fixed by http://marc.info/?l=linux-netdev&m=133287553017193&w=2

> > I have teql0 configured as a normal device, and this script
> > automatically adds the PPP devices when they come up — giving me load
> > balancing across both ADSL lines.
> >
> >
> Can I convince you to try some new sfq and sfqred stuff for load
> optimization on those lines?

Perhaps, but having instrumented the queue depths carefully for testing
the above patch, I can tell you that I almost never see more than one
packet in the uplink queue. I've only been able to stress the queueing
with 'ping -l 100'; not any realistic workload. Under normal operation,
I don't even have a second packet in the queue for the ADSL modem to
send in its TX done IRQ. It has to let the line go idle and wait for
something new to come in. Perhaps you just need to remind me what I
should be using as a test case? :)

Do you have scripts which automatically apply schedulers to interfaces
based on the network configuration (like I'm doing here for TEQL)?

-- 
dwmw2

Attachment: smime.p7s
Description: S/MIME cryptographic signature

_______________________________________________
openwrt-devel mailing list
[email protected]
https://lists.openwrt.org/mailman/listinfo/openwrt-devel

Reply via email to