Throughput testing 386.11 on RT-AC1900 (RT-AC68U)

[drinkingbird]

Had seen a couple questions around throughput and had some time to kill (and was curious) so did some testing.

TL;DR:
RT-AC1900 (Equal to RT-AC68U v1 variants with 1Ghz dual core processor) running Merlin 386.11 with extra VPNs removed.
Aiprotection and Traffic Stats enabled, but no other major throughput/CPU impacting features
Wired Switching - 930Mbit/sec
Wired Routing - 900M LAN, 880M WAN, CPU at 99%
Wireless to wired switching - 500M
Wireless to wired routing - 460M (Guest 1 to LAN), 450M (LAN or Guest 1 to WAN), CPU near pegged but not quite
Wireless to wireless "switching" - 230M

What surprised me most was the wired routing performance with aiprotection and bandwidth stats enabled. Since my internet is only 350M I'd never bothered to hook up a PC to the WAN port to truly test it. Same for inter-VLAN connectivity on the LAN (which basically is the same thing but should bypass aiprotection and some traffic stats, hence being a bit faster, but not much).

Only disappointing numbers are wireless to wireless but I was already aware of that from doing file transfers. The chipset is pretty limited in this router and wireless to wireless doubles the load on it (plus additional collisions, overhead, etc).

Test setup:
RT-AC1900 which is a 1Ghz variant of the RT-AC68U v1 (I believe same as B1/C1/E1 revisions of the 68U). Firmware is Merlin 386.11 with extra VPN NVRAM variables removed (factory reset after installing 386.11 a couple months ago).
Aiprotection enabled
Bandwidth stats enabled
No QOS or parental controls
No addons and only basic scripts which shouldn't impact throughput
Jumbo frames enabled on the switch but not on the PCs I was testing with, so all tests were 1500 MTU
Wireless clients were both Intel 2 stream 5ghz AC at 866M link speed, spaced 10 feet from the router and 10 feet from each other in a triangle
Was never able to peg the CPU (came close) on the wireless tests so my wireless environment (or the router chipset) caps out around 500M, at least at the time of this test
Both guest networks have "access intranet" disabled which means GW1 is using VLANs, different subnet, routing, EBTABLES, and IPTABLES, and GW2 is using same VLAN/subnet, routing (to WAN, didn't test LAN as that requires doing static ARPs and was too lazy), and EBTABLES+IPTABLES
Both PCs firewall/AV disabled (though they can do 1G fine with them enabled) and no CPU core ever hit 100% on either PC
Same PCs used for wired and wireless tests (placed in airplane mode for wireless tests to be sure it was going over the correct path).

I didn't watch the CPU for all tests, just ones I knew were likely CPU limited. Someday maybe I'll disable aiprotection and traffic stats to see which ones it improves, but I feel like to really test that, may have to factory reset to clear it all out.

Numbers were similar (within margin of error) for 8-stream IPERF3 and for uncompressible windows to windows file transfer. Rounded to nearest 5M for simplicity.
Throughput in both directions was also similar

Did not test full duplex (simultaneous up and down) as IPERF3 for windows doesn't have that functionality and didn't feel like digging out old iperf and adding a bunch of extra lines here. I would expect switching performance to be full speed in both directions simultaneously, and routing to be probably around half.

CONTROL
A End	B End	Passes through	Throughput	CPU	Notes
TP Link Switch Wired	TP Link Switch Wired	Switch only	950M		TP Link TL-SG108E Smart Switch
Asus LAN to LAN
A End	B End	Passes through	Throughput	CPU	Notes
Wired LAN	Wired LAN	Switch only	930M
Wired Guest 1	Wired LAN	Switch, Router, EBTABLES, IPTABLES	900M	99%​	Requires custom firewall rules / one port in VL50x
Wireless LAN	Wired LAN	Wifi Bridge, Switch	500M
Wireless Guest 1	Wired LAN	Wifi Bridge, Switch, Router, EBTABLES, IPTABLES	460M	99%​	Requires custom firewall rules
Wireless LAN	Wireless LAN	Wifi Radio Only (Double load on chipset)	230M
Wireless Guest 1	Wireless LAN	Wifi Bridge (Double load on chipset), Router, EBTABLES, IPTABLES	225M		Requires custom firewall rules
		**Unsure if AiProtection or Traffic Stats impacts any of above, in theory it shouldn't
Asus LAN to WAN
A End	B End	Passes through	Throughput	CPU	Notes
Wired LAN	Wired WAN	Switch, Router, IPTABLES, AiProtection, Traffic Stats	880M	99%​
Wired Guest 1	Wired WAN	Switch, Router, EBTABLES, IPTABLES, AiProtection, Traffic Stats	870M	99%​	Requires setting LAN port into VL50x
Wireless LAN	Wired WAN	Wifi Bridge, Router, IPTABLES, AiProtection, Traffic Stats, Switch	450M	96%​
Wireless Guest 1	Wired WAN	Wifi Bridge, Router, EBTABLES, IPTABLES, AiProtection, Traffic Stats, Switch	450M	96%​
Wireless Guest 2	Wired WAN	Wifi Bridge, Router, EBTABLES, IPTABLES, AiProtection, Traffic Stats, Switch	450M	96%​

 

[KrypteX]

Very nice work, congrats!
I have the AC68P (AC68Uv2 B1 variant) with CPU running @1GHz and I knew it can do wired NAT around 900 Mbps and WiFi NAT around half the link speed (230 on a 450 3-stream N link, or 450 on a 866 2-stream AC link).

I am curious how much the new 386.12 (still in beta) would be able to achieve. I guess it's mostly the same performance.

That being said, I'm still laughing when I still see some users having 100-300 Mbps internet connections, but complaining about this router's lack of capabilities Oh and not to mention their strong drive to buy the latest and greatest AX+ device for a lot of money. Consumerism at its best.

 

[drinkingbird]

Very nice work, congrats!
I have the AC68P (AC68Uv2 B1 variant) with CPU running @1GHz and I knew it can do wired NAT around 900 Mbps and WiFi NAT around half the link speed (230 on a 450 3-stream N link, or 450 on a 866 2-stream AC link).

I am curious how much the new 386.12 (still in beta) would be able to achieve. I guess it's mostly the same performance.

That being said, I'm still laughing when I still see some users having 100-300 Mbps internet connections, but complaining about this router's lack of capabilities Oh and not to mention their strong drive to buy the latest and greatest AX+ device for a lot of money. Consumerism at its best.

What surprised me was that it could do it with aiprotection and traffic stats on. Especially since when I speed test my normal internet at 350M the CPU is around 65 percent, but obviously there is no rule that CPU to throughput is linear. And after this test I now know watching the CPU in the GUI during a speed test actually takes up around 5% CPU. Even top/ntop uses a bit. So basically after confirming the CPU is pegged have to run the test again with the GUI closed or top exited out.

Of course attempting to do a gig worth of web browsing (would need a couple hundred people, or replay a sniffer file at high speed) would certainly cut this to a fraction as aiprotect (and the router itself) would be working much harder.

But the average home user isn't maxing out their connection on imix traffic, it's going to be on file downloads.

 

[drinkingbird]

Very nice work, congrats!
I have the AC68P (AC68Uv2 B1 variant) with CPU running @1GHz and I knew it can do wired NAT around 900 Mbps and WiFi NAT around half the link speed (230 on a 450 3-stream N link, or 450 on a 866 2-stream AC link).

I am curious how much the new 386.12 (still in beta) would be able to achieve. I guess it's mostly the same performance.

That being said, I'm still laughing when I still see some users having 100-300 Mbps internet connections, but complaining about this router's lack of capabilities Oh and not to mention their strong drive to buy the latest and greatest AX+ device for a lot of money. Consumerism at its best.

I don't think 386.12 or any new version for these routers will have any performance improvements, not of any significance anyway. The only updates to these routers at this point are security related for the most part, possibly minor bug fixes. Our routers are essentially end of life and probably destined for the end of support list soon.

I'm sure with some settings tweaks I could bump the wifi throughput up a bit, but these speeds are more than enough for my needs. I'd rather have range and reliability over a bit extra throughput.

I'm actually thinking now that when my CPUs are at 6x% CPU1 and 4x% CPU2 during a 350M internet speed test, the routing may very well be CPU2, with CPU1 being Aiprotection and traffic stats. Routing CPU usage should be pretty linear, where the other two likely are not, so that actually makes just about perfect sense, both CPUs are at 99% (ish, can't actually watch them during the fastest speed test since that consumes some CPU) when I am able to hit 880M via the WAN so that lines up approximately with CPU2 at 40% for 350M vs 100% for 880M.

For a $35 router I got on clearance at walmart like 5 or 6 years ago, it's a little workhorse. Recently my outdoor AP radio starting dying so I just unplugged it, don't really use it that much, and I was impressed to see how good the 2.4ghz range is on the Asus outside, usable anywhere in my yard, even at the neighbor's across the street, and all my outdoor cameras function very well (with aluminum siding on my house which is usually certain death for wifi). Fairly congested 2.4 band here as well.

I'm sure an AX router would get me better performance on both 5ghz and 2.4ghz but like I said, what I have now is reliable and plenty for my purposes.

 

[Tech9]

The chipset is pretty limited in this router and wireless to wireless doubles the load on it

The reduced throughput in half or more is a result of airtime sharing between the wireless clients. There is no double load on the chipset. A router with much faster CPU and better radios like RT-AC86U will produce very similar results limited by Wi-Fi technology and how Wi-Fi works in general.

 

[drinkingbird]

The reduced throughput in half or more is a result of airtime sharing between the wireless clients. There is no double load on the chipset. A router with much faster CPU and better radios like RT-AC86U will produce very similar results limited by Wi-Fi technology and how Wi-Fi works in general.

It isn't a CPU limitation so much as a wifi hardware limitation. I forget the CPU number but it was under 50% during those tests. I had a Unifi AC AP before the Asus that could do 400 to 500M between two 2 stream AC clients, wifi to wifi depending on the day and environment. The AC86U I would hope could hit 300 or more at least, but the chipset isn't really THAT much better so who knows. I have some other AC APs with better wireless chipsets I should probably toy around with but I wouldn't expect any night and day differences, if it was an issue I'd be looking at AX, AXe, or BE (but it isn't). Very rare for me to do wifi to wifi large transfers, if I need speed it is easy to plug in one (if not both) devices. Though having come from much higher rates I was a bit surprised the first time when it capped out around 250 ish.

 

[Tech9]

I had a Unifi AC AP before the Asus that could do 400 to 500M between two 2 stream AC clients

This is impossible. You have total of about 600Mbps throughput shared between all Wi-Fi clients.

 

[drinkingbird]

This is impossible. You have total of about 600Mbps throughput shared between all Wi-Fi clients.

While it is a shared medium like a hub, similar to hubs you can get much better throughput when the traffic is directly between two clients with nothing else going on, minimal collisions and timeslots can line up better. I don't recall if the Ubiquiti had any sort of throughput enhancer but it isn't something I ever would have enabled on the computers anyway. Have always noticed (since B even) that a transfer between two wireless clients can get higher throughput than two clients accessing independent things (i.e. simultaneous speed tests, simultaneous file transfers to a wired device etc).

But I definitely meant to say 300 to 400 (I specifically remember 350 being a pretty hard ceiling, I don't think I ever quite hit that, probably 330-340). The 400 to 500 was stuck in my head from the wireless to wired testing I was doing last night.

Of course that was also back in the days when tweaking your window size was still a bit of a thing. Even though Windows starting doing dynamic window sizing it doesn't yield as good of throughput on large file transfers as hardcoding it to a large size. Nowadays I don't bother and just let it do its thing.