[Cake] Bandwidith rate by host instead of global while using [dual-]srchost and [dual-]dsthost
pete at heistp.net
Sat Jun 16 15:20:22 EDT 2018
> On Jun 16, 2018, at 5:33 PM, Michel Blais <michel at targointernet.com> wrote:
> I doubt any WISP (Ubnt main market) using Edgerouter for routing would activate codel since it would hurt badly forwarding performance. Anyway, just like for CMTS, WISP bottleneck is the wireless radio, not the router.
Hi Michel- badly hurt forwarding performance in what way?
The bottleneck is not always the wireless radio, particularly when the link is not over-provisioned. I’m working with a WISP that uses Ubiquiti gear and APUs in their backhaul that do soft rate limiting (on their stable links) with HTB and SFQ. In my lab tests using NSM5s and APUs, both Cake and fq_codel outperform SFQ in both inter-flow and (especially) intra-flow latency under load, so that’s why we’re working on switching to one or the other. HTB is needed though for customized fairness, as you noted in your original question.
Out-of-the-box performance of NSM5s with airMAX enabled leaves something left to be desired as far as latency under load goes, so one should either over-provision or do some kind of shaping. Compare:
airMAX enabled, default queueing:
airMAX disabled, cake queueing (enabling airMAX here hurts inter-flow latency):
By the way, if you note the isochronous spikes in the output, that appears to be due to a bug in the NSM5’s Ethernet or internal switch driver that has been fixed in LEDE/OpenWRT. If you or anyone else happens to have an NSM5, I’d appreciate your testing and adding to the thread, where I didn’t yet manage to convince Ubiquiti that it’s real and an issue:
I’m actively writing an article on this topic and will post when finished...
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Cake