<html><head><meta http-equiv="Content-Type" content="text/html; charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Sep 7, 2018, at 1:03 AM, Jonathan Morton <<a href="mailto:chromatix99@gmail.com" class="">chromatix99@gmail.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div class=""><blockquote type="cite" class="">On 7 Sep, 2018, at 1:37 am, Pete Heist <<a href="mailto:pete@heistp.net" class="">pete@heistp.net</a>> wrote:<br class=""><br class="">This router is an old ALIX with kernel 2.6.26, but on the other hand it does have hfsc + esfq (a variant of sfq with host fairness) deployed, so if it’s actually controlling the queue, one might suspect that sfq it could control inter-flow latency at least somewhat.<br class=""></blockquote><br class="">ESFQ has two important faults: it doesn't explicitly control the length of individual queues (only tail-drops when a global limit is reached), and it suffers from hash collisions at the full "birthday problem" rate. So some of your measurement traffic is likely colliding with real traffic and suffering accordingly.<br class=""></div></div></blockquote><div><br class=""></div><div>Ah, ok, that is important.</div><br class=""><blockquote type="cite" class=""><div class=""><div class="">That still makes ESFQ far better than a dumb FIFO.<br class=""></div></div></blockquote></div><br class=""><div class="">I’ve heard tales of the way things were.</div><div class=""><br class=""></div><div class="">As a contrast, the router I’m on: <a href="https://www.heistp.net/downloads/vysina_ping.pdf" class="">https://www.heistp.net/downloads/vysina_ping.pdf</a> The big difference here is this router’s uplink is licensed spectrum full-duplex 100Mbit, whereas Jerab from earlier is 5GHz WiFi (2x NSM5). The shift around June was an upgrade from ALIX to APU.</div><div class=""><br class=""></div><div class="">I haven’t seen evidence yet of backhaul links running at saturation for long periods. When I watch throughputs in real-time I do see pulses though that probably don't show up in the long-term MRTG throughput graphs. I wonder what queue lengths look like at millisecond resolution during these events.</div></body></html>