[Cerowrt-devel] SQM in mainline openwrt, fq_codel considered for fedora default
dave.taht at gmail.com
Thu Dec 4 13:24:32 EST 2014
On Thu, Dec 4, 2014 at 8:09 AM, Tom Gundersen <teg at jklm.no> wrote:
> On Tue, Oct 21, 2014 at 10:59 PM, Dave Taht <dave.taht at gmail.com> wrote:
>> On Tue, Oct 21, 2014 at 12:51 PM, Dave Taht <dave.taht at gmail.com> wrote:
>>>> You can see that BQL makes the most difference in the latency.
>>> And ALSO that these fixes also improved system throughput enormously.
>> Meant to include that plot.
>> You can disregard the decline in download bandwidth (as we are also
>> sending 5x as many acks and measurement data, which is not counted
>> in that part of the plot)
>>> This is partially due to the improvement in ack clocking you get from
>>> reduced RTTs, partially due to improved cache behavior (shorter
>>> queues), and partially continual improvements elsewhere in the tcp
>>> portions of the stack.
>>> With more recent kernels...
>>> I now get full throughput from the beagles in both directions with the
>>> 3.16 kernel,
>>> the stil out of tree bql patch, and either fq or fq_codel. I haven't
>>> got around to plotting all those results (they are from kathie's new
>>> lab), but they are here:
>> The latency spikes are generally due to not having BQL, probably:
>> This is using the new fq scheduler on both sides, with BQL enabled.
>> The switch most likely is prioritizing EF marked packets. (as is
>> sch_fq). Most of the buffering is in the switch, not the host, now.
>> (the prior results I showed had no switch in the way)
>>> There is a small buffered tail dropping switch in the way, on these
>>> later data sets. There was some puzzling behavior on the e1000e that I
>>> need to poke into in a more controlled setting.
>>> As for other tunables on hosts, TCP small queues might be amiable to
>>> some tuning, but that too may well evolve further in kernelspace.
>> So I have now drowned you in data on one architecture and setup. The
>> most thoroughly publicly analyzed devices and drivers are the ar71xx,
>> e1000e, and beaglebone at this point.
>> The use of fq_codel in a qos system (artificially rate limited using
>> htb, hfsc, or tbf) is pretty well proven to be a huge win at this
>> At line rates fq and fq_codel still help quite a bit without a BQL
>> enabled driver, BQL is needed for best results. I don't know to what
>> extent the BQL enabled drivers already cover the marketplace, it was
>> generally my assumption that the e1000e counted for a lot...
>> And thus with all the positive results so far, more wider distribution
>> of the new stuff
>> on more devices outside the sample set of those on the bloat,
>> cerowrt-devel and codel lists (about 500 people all told),
>> and all ya gotta do is turn it on.
> Thanks for the very detailed info Dave.
> What I'm taking from this is that at the moment we should not be doing
> anything more than what we already do (the sysctl changes that Michal
> pushed), and that any improvements should be made in the kernel
> drivers (BQL support in particular).
I would like to see more pushback from everyone in the universe on new
ethernet drivers in particular, to always have BQL.
it is very useful at all speeds and there has been a big push for it
in the 10GigE+ space, but not enough at gigE and below.
So we are always asking for it when a new
driver arrives, or new work on an old driver arrives. (recently the nvneta
for example). It would be nice if there was a big push to add BQL
to older drivers (talk to jesper about that)
Recently xmit_more support arrived in the kernel mainline,
it needs BQL, and it's *very good*.
> What we could do in the future, is in case you have kernel features
> that should be enabled by default as they benefit the general user,
> but where you cannot change the kernel defaults due to backwards
> compat, we can (as we did with fq) add these to the recommended sysctl
> files shipped with systemd. This is of course just a gentle nudge for
> people to use features, distros/admins can still override them.
Cool. But I don't mind testing stuff at this level for a year or more before
inflicting it on users. :)
Relative to the kernel options, IPV6_SUBTREES is needed for good
quality source specific routing support. So far as I know that has defaulted
to on in most distros for a long time.
Relative to device naming, we are increasingly living in a world where
IP addresses come and go, and firewalls are having a lot of trouble
dealing with that - (two recent examples:
My own take on this was to name devices after their security level
and then type in cerowrt, which is not an idea that has been taken up
anywhere else - but that allows for never having to reload the firewall and
loss conntrack state on an address change or interface add, if more work was put
into it. It does look like nftables has some simpler means for adding/
deleting interfaces from a security model, but many solutions
seem to require changing a lot of state and introducing brief windows
of vulnerability or network disconnectivity.
In my case I regularly run with two or more interfaces always connected
to the internet, I never need to down one and up the other, or change
default routes, I just unplug in one place, and go to another, and
plug in again... and all my sessions stay up.... . Admittedly now that
I use mosh more often (has ipv6 support in git head), I am not
as bothered by disconnectivity as I used to be.....
Lastly I would like to see the dhcp-related work in systemd
paying attention to the kinds of issues the ipv6 homenet wg in the
ietf and openwrt (and early distributors of ipv6 native addressing,
are dealing with... and things like hnetd looked at.
>> This gets me to the stopping point that we hit a whlle back, which was
>> reliably determining if a good clocksource was present in the system.
>> Somehow. clock_getres, perhaps?
> I appear to have missed the context on this. In what situation do we
> (userspace) need to care about the existence of a high quality
There are a (very few) cpus/arches that lack a fast and high quality timesource.
way older x86 boxes lacking hpet, for example. Network performance on these
devices tends to be pretty bad in the first place, and it is simply unknown at
this point how much the per-packet timestamping in codel would change things.
Nearly all modern chips have a fast, onchip, single cycle timesource, so this
is not a problem for those.
 In fact I just noticed my ethernet was unplugged, and had been for a
d at ganesha:~/git/tinc/src$ ifconfig eth0
eth0 Link encap:Ethernet HWaddr d8:50:e6:a0:b6:32
inet addr:172.26.16.2 Bcast:172.26.16.255 Mask:255.255.255.0
inet6 addr: fe80::da50:e6ff:fea0:b632/64 Scope:Link
UP BROADCAST MULTICAST MTU:1500 Metric:1
RX packets:9508669 errors:0 dropped:0 overruns:0 frame:0
TX packets:8321719 errors:0 dropped:0 overruns:0 carrier:0
RX bytes:9637122680 (9.6 GB) TX bytes:5583512622 (5.5 GB)
d at ganesha:~/git/tinc/src$ ifconfig wlan0
wlan0 Link encap:Ethernet HWaddr 24:0a:64:cc:24:7d
inet addr:172.26.17.228 Bcast:255.255.255.255 Mask:0.0.0.0
inet6 addr: fe80::260a:64ff:fecc:247d/64 Scope:Link
inet6 addr: 2601:9:3600:b397:260a:64ff:fecc:242d/128 Scope:Global
UP BROADCAST RUNNING MULTICAST MTU:1500 Metric:1
RX packets:1743729 errors:0 dropped:0 overruns:0 frame:0
TX packets:1489781 errors:0 dropped:0 overruns:0 carrier:0
RX bytes:1077109813 (1.0 GB) TX bytes:504903872 (504.9 MB)
d at ganesha:~/git/tinc/src$ ip route
default via 172.26.17.224 dev wlan0 proto 42 onlink
188.8.131.52/22 via 172.26.17.224 dev wlan0 proto 42 onlink
169.254.0.0/16 dev eth0 scope link metric 1000
172.26.16.0/24 dev eth0 proto kernel scope link src 172.26.16.2
172.26.16.1 via 172.26.17.227 dev wlan0 proto 42 onlink
172.26.16.3 via 172.26.17.227 dev wlan0 proto 42 onlink
172.26.17.0/24 via 172.26.17.224 dev wlan0 proto 42 onlink
172.26.17.3 via 172.26.17.227 dev wlan0 proto 42 onlink
172.26.17.227 via 172.26.17.227 dev wlan0 proto 42 onlink
192.168.7.2 via 172.26.17.227 dev wlan0 proto 42 onlink
d at ganesha:~/git/tinc/src$
More information about the Cerowrt-devel