which QoS is least CPU intensive?

[consorts]

i was using lan qos on my ac3100 while on 100Mbps isp without issue.

now that my isp gave me 300Mbps, i'm finding the CPU on my ac3100
runs at 100% when i try any of the QoS schemes in my asus router.

when all types of QoS in the router are completely disabled
i get my full 300Mbps speed while the router CPU stays <50%.

assuming i can't afford to upgrade my router to a faster cpu
is there a specific QoS that is simplest and requires the least
amount of CPU effort that may offer a compromise between
lan:wan wide speed control, while not maxing out that CPU.

 

[ColinTaylor]

Use Adaptive QoS.

 

[consorts]

Use Adaptive QoS.

adaptive is actually the worst - a highest cpu user that nets the lowest usable bandwidth.

 

[ColinTaylor]

adaptive is actually the worst - a highest cpu user that nets the lowest usable bandwidth.

Strange. The #1 advantage of using Adaptive QoS is meant to be that it can use hardware acceleration and is therefore not CPU-limited like Traditional QoS is. (I assume you're not routing your traffic through a VPN or anything like that?)

 

[consorts]

not routing your traffic through a VPN or anything like that

i'm actively using 6 ssid's and one of
them is VPN dedicated thanks to yazfi.

however, it's the first core (bandwidth processing) not
the second core (vpn processing) that keeps maxing out.

 

[ColinTaylor]

In my own tests I didn't see any appreciable difference in CPU usage using any of the different queueing disciplines (SFQ, CODEL or FQ_CODEL).

however, it's the first core (bandwidth processing) not
the second core (vpn processing) that keeps maxing out.

In theory with Adaptive QoS the first core shouldn't be maxing out. Maybe it's incompatible with some third party scripts.

 

[consorts]

ok, so everyone was out of the house for a few hours, so i did a lot of trail and error and discovered that pretty much any qos setting well over 200Mbps on my ac3100 would max out the CPU, so i "settled" on the following;

under the "Manual Settings" pages, i made everything Highest 100%
so QoS would not have to think/cpu about who doing what is better,
so after all my efforts, i was always keeping the cpu under 99% use.

it's worth noting that even with QoS totally disabled, I was still getting
about "B" quality bufferbloat at my full 300:300 isp capable fios speed,
but nobody home actually needs all this speed, so I'll go for the quality
over quantity - anytime ("speed test plus" is from megapath - a clec)

here are some other threads with related discussions i considered;
https://www.snbforums.com/threads/a...wrong-or-am-i-expecting-too-much-of-it.47702/
https://www.snbforums.com/threads/qos-fios-wan-packet-overhead.40233/

one specific bit of advice for fios ont(ethernet) users is to make sure
you power up in sequence ont(wait2min) router(wait3min) then test
i was getting a lot of false results when all i did was reboot the router.

 

[RMerlin]

Back in the day, I was able to push an RT-AC56U and its 800 MHz CPU up to 650 Mbps when using Adaptive QoS (as I was testing the impact of QoS on throughput). If you are only reaching 200 Mbps, then it means you have something else disabling or bypassing hardware acceleration.

 

[consorts]

something else disabling or bypassing hardware acceleration.

thanks for showing an interest

i suspect its the vlan and vpn related to yazfi
but it's proven a very reliable app for my uses
so i'd rather keep yazfi till i get a faster a2900
than enjoy more bandwidth i don't really need.

 

[Trip]

adaptive is actually the worst - a highest cpu user that nets the lowest usable bandwidth.

Worse only in the context of throughput being CPU-bound. But for WAN links that are, say, a factor of 75% the speed of your LAN, or slower, anything using fq_codel or CAKE, properly implemented, should actually produce better packet flow behavior on egress, provided you have the CPU available to max out the WAN link. In this case, the AC3100 does not, and total throughput is being CPU-throttled.

Strange. The #1 advantage of using Adaptive QoS is meant to be that it can use hardware acceleration and is therefore not CPU-limited like Traditional QoS is. (I assume you're not routing your traffic through a VPN or anything like that?)

Queuing, and thus all of Asus's "Adaptive QoS", cannot be offloaded, so no need to go searching for more culprits behind why traffic is being CPU-driven. That's your reason right there. The only way to re-route packet flow back through Broadcom CTF (hardware acceleration) is to turn QoS off completely. That is the only way the OP will be able to get full WAN bandwidth utilization at that link speed with that model router. For anything more, if you want to run fq_codel-based queuing at more than a couple hundred Mb/s, you'll need a higher-clock ARM chip, or more ideally a MIPS architecture, or better still, an x86 box. That's the deal. Sucks, but it is what it is.

 

[ColinTaylor]

Queuing, and thus all of Asus's "Adaptive QoS", cannot be offloaded, so no need to go searching for more culprits behind why traffic is being CPU-driven. That's your reason right there. The only way to re-route packet flow back through Broadcom CTF (hardware acceleration) is to turn QoS off completely.

What you've said seems to make sense but is at odds with what @RMerlin said in post #8 and other posts. It also seems to contradict people's test results IIRC. a b

 

[Trip]

What you've said seems to make sense but is at odds with what @RMerlin said in post #8 and other posts. It also seems to contradict people's test results IIRC. a b

Well I was going to mention that the only variable here would be if Asus has a QoS-offload schema in place where they can somehow hw-accelerate this piece, but to me it seems highly unlikely that they'd be able to do it for methods that involve queuing, as it's non-offloadable in nearly every platform I'm aware of (OpenWRT -- the parent fork of most OEM firmware -- EdgeOS, RouterOS, pfSense, etc. etc.). Would love to know more about how it's done, if indeed Asus is able to offload it... @RMerlin?

 

[ColinTaylor]

I believe it's done within closed source kernel code.

https://www.snbforums.com/threads/r...and-inner-workings.36836/page-187#post-520759

https://www.snbforums.com/threads/384-9_alpha-builds-testing-all-variants.54050/page-23#post-461644

 

[sfx2000]

Well I was going to mention that the only variable here would be if Asus has a QoS-offload schema in place where they can somehow hw-accelerate this piece, but to me it seems highly unlikely that they'd be able to do it for methods that involve queuing, as it's non-offloadable in nearly every platform I'm aware of (OpenWRT -- the parent fork of most OEM firmware -- EdgeOS, RouterOS, pfSense, etc. etc.). Would love to know more about how it's done, if indeed Asus is able to offload it... @RMerlin?

Broadcom and QC-Atheros both have interesting off-loads that are close source. Discussion here is with Broadcom, but QC-Atheros does have very interesting things on the IPQ series SoC's

OpenWRT has been doing some effort on software flow offload with the netfilter team - mixed result. as switch documentation is hard to find...

 

[sfx2000]

And for what it is worth - I'm doing work on AR9331, and there, SQM in SW with Cake works fine, but not a lot of demand with clients there.

Use Profile is different here - it's IoT, so BW is low demand, but latency important....