Just thought I would share recent findings after testing SQM implementation along with software flow offloading on the R1 Plus LTS running v23.05.2.
I have a 1000/20 mbps cable service from Xfinity. After implementing SQM a couple months ago and being happy with a notable improvement to buffer bloat tests, I let it be ...
I decided to look at overall speed test and latency performance and it was evident that using cake was significantly reducing download speeds while only adding marginal latency gains in comparison to fq_codel. For example:
- flow offload disabled: ~550 mpbs
- flow offload enabled: ~850 mbps sustained (bursts up to a gig)
- SQM enabled using cake, flow offloading enabled: ~350 mbps
- SQM enabled using fq_codel, flow offloading enabled: ~750 mbps
Cake has slightly better latency than fq_codel, ~5ms lower download under load, and no difference on upload latency.
I've set my ingress/egress values to 700000/20000.
I have not researched in depth the underlying technology of SQM queuing, yet am happy with better overall throughput combined with nearly identical latency using fq_codel. While everything I've read points to cake as being better, I am curious to know why it cuts download bandwidth by 50% ...?