<div dir="ltr"><div class="gmail_default" style="font-family:verdana,sans-serif">These are just the fq settings as they get applied when having fq as default qdiscs. I guess there are room for improvements on those default settings depending on use case.</div><div class="gmail_default" style="font-family:verdana,sans-serif"><br></div><div class="gmail_default" style="font-family:verdana,sans-serif">For future reference: should I increase the limit on drops or is it okay as it is?</div></div><div class="gmail_extra"><br><div class="gmail_quote">On 26 January 2017 at 00:53, Eric Dumazet <span dir="ltr"><<a href="mailto:eric.dumazet@gmail.com" target="_blank">eric.dumazet@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class="">On Thu, 2017-01-26 at 00:47 +0100, Hans-Kristian Bakke wrote:<br>
><br>
><br>
> I did record the qdisc settings, but I didn't capture the stats, but<br>
> throttling is definitively active when I watch the tc -s stats in<br>
> realtime when testing (looking at tun1)<br>
><br>
><br>
> ​tc -s qdisc show<br>
> qdisc noqueue 0: dev lo root refcnt 2<br>
>  Sent 0 bytes 0 pkt (dropped 0, overlimits 0 requeues 0)<br>
>  backlog 0b 0p requeues 0<br>
> qdisc fq 8007: dev eth0 root refcnt 2 limit 10000p flow_limit 100p<br>
> buckets 1024 orphan_mask 1023 quantum 3028 initial_quantum 15140<br>
> refill_delay 40.0ms<br>
>  Sent 1420855729 bytes 969198 pkt (dropped 134, overlimits 0 requeues<br>
> 0)<br>
>  backlog 0b 0p requeues 0<br>
>   124 flows (123 inactive, 0 throttled)<br>
>   0 gc, 0 highprio, 3 throttled, 3925 ns latency, 134 flows_plimit<br>
<br>
</span>You seem to hit the "flow_limit 100" maybe because all packets are going<br>
through a single encap flow. ( 134 drops )<br>
<span class=""><br>
> qdisc fq 8008: dev tun1 root refcnt 2 limit 10000p flow_limit 100p<br>
> buckets 1024 orphan_mask 1023 quantum 3000 initial_quantum 15000<br>
> refill_delay 40.0ms<br>
>  Sent 1031289740 bytes 741181 pkt (dropped 0, overlimits 0 requeues 0)<br>
>  backlog 101616b 3p requeues 0<br>
>   16 flows (15 inactive, 1 throttled), next packet delay 351937 ns<br>
>   0 gc, 0 highprio, 58377 throttled, 12761 ns latency<br>
> ​<br>
><br>
<br>
</span>Looks good, although latency seems a bit high, thanks !<br>
<div class="HOEnZb"><div class="h5">><br>
><br>
> On 26 January 2017 at 00:33, Eric Dumazet <<a href="mailto:eric.dumazet@gmail.com">eric.dumazet@gmail.com</a>><br>
> wrote:<br>
><br>
>         On Thu, 2017-01-26 at 00:04 +0100, Hans-Kristian Bakke wrote:<br>
>         > I can do that. I guess I should do the capture from tun1 as<br>
>         that is<br>
>         > the place that the tcp-traffic is visible? My non-virtual<br>
>         nic is only<br>
>         > seeing OpenVPN encapsulated UDP-traffic.<br>
>         ><br>
><br>
>         But is FQ installed at the point TCP sockets are ?<br>
><br>
>         You should give us "tc -s qdisc show xxx"  so that we can<br>
>         check if<br>
>         pacing (throttling) actually happens.<br>
><br>
><br>
>         > On 25 January 2017 at 23:48, Neal Cardwell<br>
>         <<a href="mailto:ncardwell@google.com">ncardwell@google.com</a>><br>
>         > wrote:<br>
>         >         On Wed, Jan 25, 2017 at 5:38 PM, Hans-Kristian Bakke<br>
>         >         <<a href="mailto:hkbakke@gmail.com">hkbakke@gmail.com</a>> wrote:<br>
>         >                 Actually.. the 1-4 mbit/s results with fq<br>
>         sporadically<br>
>         >                 appears again as I keep testing but it is<br>
>         most likely<br>
>         >                 caused by all the unknowns between me an my<br>
>         >                 testserver. But still, changing to<br>
>         pfifo_qdisc seems<br>
>         >                 to normalize the throughput again with BBR,<br>
>         could this<br>
>         >                 be one of those times where BBR and pacing<br>
>         actually is<br>
>         >                 getting hurt for playing nice in some very<br>
>         variable<br>
>         >                 bottleneck on the way?<br>
>         ><br>
>         ><br>
>         >         Possibly. Would you be able to take a tcpdump trace<br>
>         of each<br>
>         >         trial (headers only would be ideal), and post on a<br>
>         web site<br>
>         >         somewhere a pcap trace for one of the slow trials?<br>
>         ><br>
>         ><br>
>         >         For example:<br>
>         ><br>
>         ><br>
>         >            tcpdump -n -w /tmp/out.pcap -s 120 -i eth0 -c<br>
>         1000000 &<br>
>         ><br>
>         ><br>
>         ><br>
>         >         thanks,<br>
>         >         neal<br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
>         >                 On 25 January 2017 at 23:01, Neal Cardwell<br>
>         >                 <<a href="mailto:ncardwell@google.com">ncardwell@google.com</a>> wrote:<br>
>         >                         On Wed, Jan 25, 2017 at 3:54 PM,<br>
>         Hans-Kristian<br>
>         >                         Bakke <<a href="mailto:hkbakke@gmail.com">hkbakke@gmail.com</a>> wrote:<br>
>         >                                 Hi<br>
>         ><br>
>         ><br>
>         >                                 Kernel 4.9 finally landed in<br>
>         Debian<br>
>         >                                 testing so I could finally<br>
>         test BBR in<br>
>         >                                 a real life environment that<br>
>         I have<br>
>         >                                 struggled with getting any<br>
>         kind of<br>
>         >                                 performance out of.<br>
>         ><br>
>         ><br>
>         >                                 The challenge at hand is UDP<br>
>         based<br>
>         >                                 OpenVPN through europe at<br>
>         around 35 ms<br>
>         >                                 rtt to my VPN-provider with<br>
>         plenty of<br>
>         >                                 available bandwith available<br>
>         in both<br>
>         >                                 ends and everything<br>
>         completely unknown<br>
>         >                                 in between. After tuning the<br>
>         >                                 UDP-buffers up to make room<br>
>         for my 500<br>
>         >                                 mbit/s symmetrical bandwith<br>
>         at 35 ms<br>
>         >                                 the download part seemed to<br>
>         work<br>
>         >                                 nicely at an unreliable 150<br>
>         to 300<br>
>         >                                 mbit/s, while the upload was<br>
>         stuck at<br>
>         >                                 30 to 60 mbit/s.<br>
>         ><br>
>         ><br>
>         >                                 Just by activating BBR the<br>
>         bandwith<br>
>         >                                 instantly shot up to around<br>
>         150 mbit/s<br>
>         >                                 using a fat tcp test to a<br>
>         public<br>
>         >                                 iperf3 server located near<br>
>         my VPN exit<br>
>         >                                 point in the Netherlands.<br>
>         Replace BBR<br>
>         >                                 with qubic again and the<br>
>         performance<br>
>         >                                 is once again all over the<br>
>         place<br>
>         >                                 ranging from very bad to<br>
>         bad, but<br>
>         >                                 never better than 1/3 of<br>
>         BBRs "steady<br>
>         >                                 state". In other words<br>
>         "instant WIN!"<br>
>         ><br>
>         ><br>
>         >                         Glad to hear it. Thanks for the test<br>
>         report!<br>
>         ><br>
>         >                                 However, seeing the<br>
>         requirement of fq<br>
>         >                                 and pacing for BBR and<br>
>         noticing that I<br>
>         >                                 am running pfifo_fast within<br>
>         a VM with<br>
>         >                                 virtio NIC on a Proxmox VE<br>
>         host with<br>
>         >                                 fq_codel on all physical<br>
>         interfaces, I<br>
>         >                                 was surprised to see that it<br>
>         worked so<br>
>         >                                 well.<br>
>         >                                 I then replaced pfifo_fast<br>
>         with fq and<br>
>         >                                 the performance went right<br>
>         down to<br>
>         >                                 only 1-4 mbit/s from around<br>
>         150<br>
>         >                                 mbit/s. Removing the fq<br>
>         again regained<br>
>         >                                 the performance at once.<br>
>         ><br>
>         ><br>
>         >                                 I have got some questions to<br>
>         you guys<br>
>         >                                 that know a lot more than me<br>
>         about<br>
>         >                                 these things:<br>
>         >                                 1. Do fq (and fq_codel) even<br>
>         work<br>
>         >                                 reliably in a VM? What is<br>
>         the best<br>
>         >                                 choice for default qdisc to<br>
>         use in a<br>
>         >                                 VM in general?<br>
>         ><br>
>         ><br>
>         >                         Eric covered this one. We are not<br>
>         aware of<br>
>         >                         specific issues with fq in VM<br>
>         environments.<br>
>         >                         And  we have tested that fq works<br>
>         sufficiently<br>
>         >                         well on Google Cloud VMs.<br>
>         ><br>
>         >                                 2. Why do BBR immediately<br>
>         "fix" all my<br>
>         >                                 issues with upload through<br>
>         that<br>
>         >                                 "unreliable" big BDP link<br>
>         with<br>
>         >                                 pfifo_fast when fq pacing is<br>
>         a<br>
>         >                                 requirement?<br>
>         ><br>
>         ><br>
>         >                         For BBR, pacing is part of the<br>
>         design in order<br>
>         >                         to make BBR more "gentle" in terms<br>
>         of the rate<br>
>         >                         at which it sends, in order to put<br>
>         less<br>
>         >                         pressure on buffers and keep packet<br>
>         loss<br>
>         >                         lower. This is particularly<br>
>         important when a<br>
>         >                         BBR flow is restarting from idle. In<br>
>         this case<br>
>         >                         BBR starts with a full cwnd, and it<br>
>         counts on<br>
>         >                         pacing to pace out the packets at<br>
>         the<br>
>         >                         estimated bandwidth, so that the<br>
>         queue can<br>
>         >                         stay relatively short and yet the<br>
>         pipe can be<br>
>         >                         filled immediately.<br>
>         ><br>
>         ><br>
>         >                         Running BBR without pacing makes BBR<br>
>         more<br>
>         >                         aggressive, particularly in<br>
>         restarting from<br>
>         >                         idle, but also in the steady state,<br>
>         where BBR<br>
>         >                         tries to use pacing to keep the<br>
>         queue short.<br>
>         ><br>
>         ><br>
>         >                         For bulk transfer tests with one<br>
>         flow, running<br>
>         >                         BBR without pacing will likely cause<br>
>         higher<br>
>         >                         queues and loss rates at the<br>
>         bottleneck, which<br>
>         >                         may negatively impact other traffic<br>
>         sharing<br>
>         >                         that bottleneck.<br>
>         ><br>
>         >                                 3. Could fq_codel on the<br>
>         physical host<br>
>         >                                 be the reason that it still<br>
>         works?<br>
>         ><br>
>         ><br>
>         >                         Nope, fq_codel does not implement<br>
>         pacing.<br>
>         ><br>
>         >                                 4. Do BBR _only_ work with<br>
>         fq pacing<br>
>         >                                 or could fq_codel be used as<br>
>         a<br>
>         >                                 replacement?<br>
>         ><br>
>         ><br>
>         >                         Nope, BBR needs pacing to work<br>
>         correctly, and<br>
>         >                         currently fq is the only Linux qdisc<br>
>         that<br>
>         >                         implements pacing.<br>
>         ><br>
>         >                                 5. Is BBR perhaps modified<br>
>         to do the<br>
>         >                                 right thing without having<br>
>         to change<br>
>         >                                 the qdisc in the current<br>
>         kernel 4.9?<br>
>         ><br>
>         ><br>
>         >                         Nope. Linux 4.9 contains the initial<br>
>         public<br>
>         >                         release of BBR from September 2016.<br>
>         And there<br>
>         >                         have been no code changes since then<br>
>         (just<br>
>         >                         expanded comments).<br>
>         ><br>
>         ><br>
>         >                         Thanks for the test report!<br>
>         ><br>
>         ><br>
>         >                         neal<br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
>         ><br>
><br>
>         > ______________________________<wbr>_________________<br>
>         > Bloat mailing list<br>
>         > <a href="mailto:Bloat@lists.bufferbloat.net">Bloat@lists.bufferbloat.net</a><br>
>         > <a href="https://lists.bufferbloat.net/listinfo/bloat" rel="noreferrer" target="_blank">https://lists.bufferbloat.net/<wbr>listinfo/bloat</a><br>
><br>
><br>
><br>
><br>
><br>
<br>
<br>
</div></div></blockquote></div><br></div>