Its been a bit of a steep learning curve but I am very slowly getting there:)
I am using openwrt as a 'bump in the wire' between a 5g router (in ip passthrough mode) and a deco x20 mesh system.
I want to use the openwrt box for SQM- actually cake -autorate- to help stabilize and reduce bufferbloat on the 5g mobile broadband
The box is a shuttle xpc dh100 with an intel i3 6100 skylake running at 3.7ghz. it has two intel lan ports.
Download speed without any filtering is 320mb/s, upload 30mb/s.
I have checked the default boxes on the luci sqm page- i.e cake, and piece of cake. I set download to 300000 kbit/s, and upload to 25000kbit/s.
In the cake-autorate config file i set min download to 40000mb/s, base to 300000mb/s, and max to 320000mb/s
min upload is 10000mb/s, base upload to 25000mb/s, and max to 30000mb/s.
Autorate is enabled and started as a service.
BUT- with cake-autorate, my download speed drops to 150-170mb/s. upload seems less affected- 20mb/s.
When i bought this shuttle i did quite a bit of research, and I always got positive results about the cpu being easily capable of 300-400mb/s with SQM enabled. In fact, most replies say it is capable of a lot more faster speeds.
running htop when doing a speedtest gives less than 5% on each thread, so a max of 20% cpu usage.
In bios I disabled c states, and EIST.
I have enabled packet steering, software flow offloading, and also irq balance.
makes no difference.
bufferbloat gives grade A
Surprisingly-for a 5g NRSA connection- i get max 2 ms on upload, soo.....i set upload on sqm and cake-autorate to '0' in case that would help download,but that didnt help download at all.
I seem to get conficting responses about sqm on a transparent bridge, ( i have this set up as a bump in the wire) saying 'Ifb' can affect the bandwidth a lot, but a 50% drop is even more than that would account for.
Make sure to enable "Detailed CPU time..." in htop's set up otherwise it will not show softinterrupt /SIRQ) context, but that is where cake spends most of its cycles...
Other than that, your latency targets might be to tight to allow higher rates? Mobile scheduling often wants to see a queue build up before it assigns more transmit opportunities, so on Mobile there might be a different throughput latency tradeoff required than on fixed networks.
Thanks very much for the reply.
Using htop, with detailed cpu, the magenta ( sirq ) 'bars' never go above 5%. The total cpu usage is about 15%.
THe shuttle PC I am using uses two intel lan ports, I211, and I1219LM, which should be perfectly capable.
I have enabled packet steering, irqbalance, software offloading, I have an ethernet overhead of 44 on the link layer adaption page. Incoming( non SQM ) is 320mb/s. With cake sqm + autorate it is 175mb/s !! I have no idea why that is??
OK, so, let's assume then you have plenty of CPU cycles, time to look at the cake-autorate configuration... maybe your latency targets are too tight... I have been told (but have not independently confirmed) that mobile carriers sometimes/often? use the amount of queued data in uplink direction as "driver" for assigning upload slots, so it seems possible to get into a regime where the carrier throttles harder than necessary if the set latency goals are so tight that the modem never sees sufficient upload queued data to keep getting assigned upload slots at a constant enough rate.
I note that cake-autorate can/does by default? generate logs that often can help to understand what is happening a bit better than just operating on first principles, maybe collect a log with a capacity test from bufferbloat.libreqos.com then post the result (first expand the Advanced Details section, then click "Download JPG" in the Share results section) as well as a link to the autorate logfile...