This one is wildly inaccurate. I ran it three times and I got grades of C, B, and A. It doesn’t come close to saturating my connection, peaking only around 150Mbs, so not sure how it can purport to measure bufferfloat.
And because it doesn’t strain my connection, it uses less CPU power, too.
Depends on the network distance,
Agreed. I am that odd edge case. I’ll fill the outbound pipe while streaming and those streaming platforms are awfully picking about receiving ACKs on their windows or I get the dreaded “buffering” or “Something went wrong” messages.
Nope. It’ll keep up with no issues. R6S with Quad-core ARM Cortex-A76(up to 2.4GHz) and quad-core Cortex-A55 CPU (up to 1.8GHz). Eight cores doesn’t hurt at all. SQM Cake→Piece of Cake uses about 75% of a single A76 core at 1Gb/s and 100% of two A76 cores at 2.5Gb/s. The latter requires mucking around with CPU affinity to ensure you don’t get stuck on a single core or worse on the A55 cores as it will use all 4 and limit throughput to about 800 Mb/s. I used htop for testing so I wasn’t fooled by sampling times in Luci. Luci is certain that total utilization never exceeds 15% under any circumstances. LoL
It is a shame you don’t have g-fiber. 1Gb→$70, 3Gb→$100, and 8Gb→$150/mo. In two years we’ve had 37 minutes of down time and two credits that exceeded expectations.
So you’re probably wondering why change at this point… VLANS! The R6S has an issue with VLANs on the 1Gb/s port. So when the two LAN ports are bridged (eth0 1Gb, eth2 2.5Gb) and DSA is used to establish VLANs on eth2, DSA puts eth0 into VLAN filter mode. The GMAC has a bug that doesn’t appreciate that and it immediately kills all IO to eth0. I’m guessing the bridge detects the GMAC on eth0 has stopped sending packets and goes into an error state. Packets stop falling out of eth2 as well. Removing eth0 from the bridge effectively resolves the issue and I’m back in business, albeit with only one LAN port. Due to inter-VLAN routing, I end up with a bottleneck. It’s time to upgrade. Now that I’m looking hard at this, I might just go with a 10GbE and upgrade to 8Gb/s service.
What packages are you using for your graphs? My experience has been that most of the graphs use averages instead of peaks. For instance, htop shows 1Gb/s SQM as 75% of a single A76 core. Meanwhile Luci is sure that 15% is as bad as it gets.
At some point I saw a write up (I think in this forum, haven’t found it again) that showed how many processor cycles it took to process SQM for a specific speed. For example, let’s say 12000 CPU cycles for 1Gb/s… don’t quote that, I made those numbers up. So with a number of CPU cycles per bits, one can size the CPU to meet the connection’s need. One can also look at the ability for specific SQM discs to thread so you can know if the you have to size for a single thread or can safely use multiple cores/threads. Thus my question.
I’d highly recommend installing htop and running those tests again. You might be surprised.
Looks like their hosts are not well peered. Live traceroute is showing bottlenecks and high pings. Since reliable pings are important for bufferbloat tests, there is no way they can get an accurate picture. Depending on where your ISP peers, you might end up close enough or have no bottlenecks between you and them.
pawl,
I just wanted to say thanks. I looked at the XGS1210-12 and it looks nice and well priced. Then I looked at the TOH and page to see how OpenWRT support is coming along. There was a mention of the XGS1250-12 3x10GbE, 1xSFP+, 8x1GbE and the same SoC. I can suffer the Zyxel web interface until OpenWRT support is a bit more stable. Knowing that it will come along and everything can have OpenWRT makes me a happy camper. Thank you!
Now I just need to settle on the router component.
1G/500M for €70 p/m is the fastest that most ISP offer here in Germany (in the foreseeable future). There's one ISP who offers a 2,5G/1,25G line for €110 p/m.
In e.g. Switzerland you could get 10G/10G for less than €45.
What a bummer...i guess nobody ever had the idea to test this thoroughly.
Not only in Switzerland ...
Router: similar to the HP T740 there is also Lenovo M920q 8500T with additional riser card and PCIe NIC. Compared to the hp t740, the m920q is smaller and has coreboot support. It is pricier though.
About your R6S and SQM: did you also test with fq_codel instead of cake?
(fq_codel should be less cpu hungry)
This if you are in US (it is meant to be workstation if you dont look at networking)
I didn’t, but the cool thing is, it isn’t hard to do…
Looks like 68% of a single A76 core using fq_codel and simplest disc. Oddly I drop about 60Mb/s in upload speed using fq_codel.
I managed to find an X550-AT2 based card on ebay that supported Multi-Gig for $20. It has an HP part number so I’m hopeful.
Nice, connect and go….
840138-001 / 817745-B21 can easily be found below $20, if that's what you bought, but I have no idea how well they fit in a vanilla PC.
If it doesn’t my HP Microserver Gen 8 might get an upgrade. Part number is 817743-001. HP’s “data sheet” has absolutely no data on it and references the wrong part number. For $20 it’s a cheep lesson if it doesn’t work.
Because HP and HPE are different companies… Yours is scavenged from datacenter.
Not really, I've been using 1Gb/s+ internet for many years and never find any need with QoS, so you also make me curious what's the actual use case that makes it very important on such a high speed line?
That's quite expensive for many people.....
But when it goes up, is QoS still needed?
The nearest commonly known application might be stock trading. Lots of small queries that are time sensitive. Add to that constant large multi-gigabyte transfers and TV streaming. Getting ACKs prioritized ensures it all runs smoothly.
My life expectancy is greatly improved when buffering errors don’t interrupt Wednesday Adams’ furthering her role as an outcast or my housemate’s TikTok addiction.
(24 hours) * 3 Gbps = 32.4 terabytes uploaded. Yeah...