is just a matter of running a command from netconsole. heatsink may not even be needed. easier than recovering your brick via serial
Okay, you are right, these are not the last 5 test results, but randomly sampled from (as far as I can tell the last 400000 speedtests). I happen to see 5 results per ISP though, instead of the 4 you reported (https://www.dslreports.com/speedtest/results/bufferbloat):
Com Hem Bredband Fiber +42ms +10ms A+ A+ A+ A+ C
Since you only reported 4 values, what was the fifth when you looked? Also note the last 400K tests only contained (right now https://www.dslreports.com/speedtest/results/isp/r2845-com-hem-bredband?further=1):
From the last 400k tests
* There are 208 results in this sample from 44 unique IPs
[...]
Buffer Bloat Grades
A+ .. 58% ..
A .. 22% ..
B .. 7% ..
C .. 9% ..
D .. 0% ..
F .. 4% ..
In my book that is still a small sample/small N (especially the 44 unique IPs), feel free to disagree.
P.S.: I retract my promise to let it go unilaterally, and reserve the right to chime in (it is a public forum, and I believe the discussion is still civil enough)
So the way forward is a separate x86-64 router, separate switch and separate AP?
Unless you want to go with mvebu (e.g. Linksys WRT3200ACM/ WRT32xx), yes - both will work for 1 GBit/s WAN speeds, x86_64 provides more headroom for future needs.
For 500Mbps or more i think that's absolutely the right way forward. I'd also recommend a smart switch that handles LAG groups and DSCP based QoS. At high speeds the amount of processing that can be done per packet is less, but also stupid algorithms like weighted round robin do ok when you have a gigE connection.
Bufferbloat and packet loss are still definitely an issue well into the 500-1000 Mbps range if you do VOIP or game traffic or video surveillance or stuff like that. When you have a gigE WAN the slowest link can easily be on your LAN such as a Wi-Fi AP or a power line modem or the like and you can need to handle SQM type stuff on these slow LAN links
1 Like
Which isn't true.... just a very quick example
Shopping for home router (2 days ago for instance)
It makes little sense to buy MIPS today since pricing is more or less the same and performance is much lower (both processing and also usually flash/memory) however it might be a good option in some scenarios (long distance links etc). It's also something the majority of all vendors are moving away from so in terms of longetivity and having the performance in mind ARM platforms are usually a better choice in terms of theoretical longetivity.
Please suggest something else you feel is appropriate
Edit: Less heated....
Yes, that would probably best solution but also the most expensive.
1 Like
SQM not so much but for me mwan3 not compatible with flow-offload has been a deal-breaker.
IMO at this point Celeron J1900 based box is the best balance of performance, cost and world-wide availability.
…and at least 2 ethernet interfaces are easy to accomplish, with the help of a 5 buck PCIe 1 GBit/s ethernet card.
A well selected (Atom based) x86_64 system can get along with as little as ~6 watts idle power consumption - a non-Atom core-i may be fine tuned to an idle power consumption of ~11-15 watts. Modern ARMv7/ ARMv8 based high-end routers are often scratching the 20 watts mark.
seems like the efficiency-gap is widening 
i7-laptop (32gig-ddr4) here uses about 6watts when viewing fhd youtube content over wifi... 3W idle.
1 Like
What are you measuring, only the power usage of one CPU core alone for this task?
powertop claims this is for the whole shebang
What does a Wattmeter show in comparison to powertop?
good call!
seems the 120W psu only has a ~60% efficiency, so powermeter shows 5,3W idle at the wall. (and some 12W with the e-dock)
1 Like
I agree.
At home I'm running with a Celeron J1900 x86-64 as Router + 8 Port GB. Switch + Archer C2600 as dumb AP and I have no performance problems at all (600 MB. of fiber connection, not 1 Gb., but is should work in 1 GB. indeed).
my Archer c7 with offload randomly crashed (a lot!) so I ended replacing it with Uqiquity ER-4 router that can process 3,4m packets and process 1Gbit traffic while using less 20% of its 4-core MIPS CPU.
I still use it as OpenWRT AP and it works great.
Just a quick note on the J1900. While I've used them for many years, I've used them for many years. There are better CPUs in the 10-W TDP class available today. There are also well-known problems with the BIOS (ACPI, in particular) with many of the J1900-based boards.
As one example, with Passmark CPU and single-thread as one notion of performance
- J4150 -- 2634 -- 1070
- J1900 -- 1839 -- 535
4 Likes
The Valleyview (Baytrail-D, e.g. Pentium J1900) ACPI problems, apparently due to a silicon bug resulting in the CPU freezing up if punit sideband communication to the integrated GPU happens while the CPU is in a c6/ c7 cstate, are indeed a major issue with this Atom generation (performance and power consumption are otherwise pretty good). It is not fun if your system locks up regularly, somewhere between every 30 minutes (kernel 5.2) or daily, weekly, monthly, … Fortunately the root cause seems to have finally been identified, with a patch set hopefully working around the silicon issues going into kernel 5.3 (since applying it, my J1900 hasn't crashed in drumroll four whole days…, knocking on wood)...
That said, my idle power consumption figures above have been measured with a good quality watt meter at the wall, taking the complete system and PSU into account.
1 Like
Mine is rock-solid, though running a non-Linux kernel. However, the ACPI apparently reports incorrectly that it has no VGA, causing it to fail to boot with said non-Linux kernel either under legacy or UEFI boot when the system console tries to load. Workaround is to set kern.vty=sc in /boot/loader.conf
Yes, measuring power draw at the wall is most appropriate. Many PSUs, even the "Gold" and "Platinum" ones, are very inefficient at low power, and finding a good one below 500 W or so has become challenging (only need ~40W with the box running a full-bore, all-core OS build). I use a high-efficiency, 12 V "brick" and a picoPSU these days -- between $0.30/kWh and being totally silent, the cost is worth it to me.
Yes, that was/ is my intention as well, except that I avoided the picoPSU and went with a board with DC input (ASRock Q1900DC-ITX), which saves a little more power, and used a good spare 19V notebook PSU (plus a 2.5" HDD and 8 GB RAM).
(system (UEFI) running linux v5.2.2 (respectively the current linux-stable or linus git HEAD of the day, VGA, DVI all fine - hopefully the recurrent/ irregular crashes are finally solved. In case of FreeBSD it may be possible that the kernel never really enters c6/ c7 cstates, I could avoid my 30-minute crashes by running md5sum /dev/zero under screen…).
1 Like