Poor performance from SQM at 350mbps (WRT3200ACM)

@dlakelan

So the plot thickens. I spent quite a few hours the last few days setting up SQM on a seperate x86 box (dual core 2.9Ghz). Only to find... the exact same behaviour. :-/

Running Modem->x86 Box->PC on 1gig ethernet directly over Cat6 cables still ended me up with erratic downloads speeds capping out around 250mbps when SQM was enabled at 330mbps. WIth SQM disabled I got the full 350mbps WAN speed.

I will redo the test tomorrow and send you the log files from your script.

I do have a theory... the cable connection I have from Virgin Media is well known to be highly speed variable during peak hours. On their forums people regularly report that during peak evening times their speed (on an advertised 350mbps package) can drop as low as 50mbps.

If the speed of the SQM managed link is varying widely and dropping well below the configured ingress speed - would that have a dramatically adverse effect on SQM performance?

Yes, it would. I have some custom scripts that might work better under these conditions. will be a few days before I can get them to you.

EDIT: by the way, you will not regret getting that x86 router up and running. so don't worry that you've wasted your time.

yea, welcome to "high-speed" cable.
if your isp is the bottleneck, no sqm-setup on your end will make any difference.
i "downgraded" from 200mbps to 50 three years ago for exactly this reason.

rating internet connection for capacity is like rating a camera for raw resolution...

Yes, ingress shaping only works if the shaper has control over the true bottleneck (or rather if the shaper is the artificial bottleneck) otherwise the bottleneck will move back into the CMTS and bufferbloat will be as without enabling SQM. Now if your ISP allows its CMTS-shaper to use burst of above-contracted bandwidth to make up for more sparse transmit opportunities on a congested link, this might make cake think that it needs to throttle aggressively to get back to the desired ingress rate and that would manifest as a relatively high bandwidth loss with cake as compared to without. In that case it might make sense to try to set cake to a larger interval/target value so it gets a bit more tolerant to bursting (but note that will also increase the latency under load increase, so this does not come for free)

Indeed. I have run a few more tests today in 'off-peak' hours and SQM is running just fine at 330mbps on the x86 box.

During peak hours with SQM disabled the throughput can vary during speed tests, but not by much and stays fairly consistent through the test. With SQM enabled however, when the line speed drops below the SQM rate threshold the impact is much more severe and results in huge spikes in latency and I suspect some dropped packets - so much so it causes DSL Reports to abort the speedtest in many cases, which is something I've never seen before.

Does it do the same with fq_codel instead of cake?