On Mon, Sep 15, 2014 at 8:27 PM, Andy Furniss <adf.lists@xxxxxxxxx> wrote: > Dave Taht wrote: > >>> tc qdisc add dev ifb0 parent 1:1 handle 10: fq_codel noecn target >>> 50ms interval 100ms >> >> >> the fq_codel internet draft[1] suggests that "interval" >> >> SHOULD be set on the order of the worst-case RTT through the >> bottleneck to give end-points sufficient time to react. >> >> so if you are using a satellite comms system, use 1200ms as the >> interval, if you are in new zealand, use 240ms, for most of the rest >> of the world, 100ms is the default and "good enough". >> >> As for "interval" Thinko: meant "target" >> >> The default target value is 5 ms, but this value SHOULD be tuned to >> be at least the transmission time of a single MTU-sized packet at >> the prevalent egress link speed (which for e.g. 1Mbps and MTU 1500 >> is ~15ms). >> >> and should be 5-10% of the interval, regardless. We hope to remove >> the target parameter entirely in some future version of codel - it's >> actually the interaction with tbf/htb/hfsc which buffers up an extra >> packet here that requires it be increased at very low bandwidths. >> >> So in this case, 2mbit, a target 8ms would be sufficient. It doesn't >> hurt to go bigger, but 50ms is overmuch. > > > Thanks for that, I am really out of date with the new stuff. > > While doing a quick test of this I just noticed that fq_codel saw > maxpacket 4542 but htb didn't count a giant - so I guess htbs limit is > higher than it used to be at least. > > Of course the packet was fake (I know nothing that big goes on the > wire), but it serves to remind me that generic-receive-offload is on by > default and if I were to shape for real on this box it would be better > off (for low speed shaping). > > ethtool -K eth0 gro off Note that on a router GRO should be off on *all* interfaces. In fact I recommend tso,gso,gro, and lro all be off on a router. Let packets be packets! There is a debloat.sh script out there that helps. See also: http://www.bufferbloat.net/projects/codel/wiki/Best_practices_for_benchmarking_Codel_and_FQ_Codel If anyone out there is up for patching new kernels and doing some benchmarking, we have a new all-singing-all-dancing "tbf + diffserv + fq_codel" single line rate limiting qdisc under development, tenatively called "cake". If you want a shot at it, let me know. > > -- Dave Täht https://www.bufferbloat.net/projects/make-wifi-fast -- To unsubscribe from this list: send the line "unsubscribe lartc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
