Gigabit PPPoE and Intel Drivers
-
@thegriffin said in Gigabit PPPoE and Intel Drivers:
Celeron 3865U
The single thread rating of that is higher than the E4500 so if it scales directly I would expect it to be fine. I have never tested one though. All CPUs are affected.
Steve
-
Thanks for the input Steve, then I'll stick to that one.
-
This post is deleted! -
Hi Tony, thanks for your post. Yes an 8th gen i5 is a lot of power, maybe you could use it for something else too with virtualization.
I'm probably going for the 3865U Celeron as it would be fine for my use, only this PPPoE issue was making me doubt but now I'm sure it'll be good.
I could get an i3-7100U regardless, though possibly a good chunk the 70 euros difference is for the marketing, but then for only 30 euros more I could get an i5-7200U and at that point I'll have upsold to myself once more
-
Hi guys, I've found another box (Partaker) that has Intel 82583V NICs (em driver) instead of the Intel I211-AT (igb driver), per this thread the em driver mitigates the PPPoE issue.
The box also has a few other advantages and I'm about to pull the trigger, my only doubt is those NICs are older (discontinued Q1 2020 vs. 1H 2029 planned for the I211-AT).
In your opinion what's the risk that they'll become unsupported in the not too distant future? I'm not that concerned about FreeBSD, more about ESXi in case I want to use virtualization in future (not planning to for now).
Thanks for sharing your thoughts.
EDIT: the processor would be either a Celeron 3865U or an i3-7100U and as per the above posts either should manage wire speed regardless of the NICs driver, still if the em driver means less CPU effort all the better and as I mentioned the Partaker box has other advantages over the Qotom and other boxes with the I211-AT NICs.
-
The PPPoE issue affects all drivers/NICs AFIAK.
The issue is that it can only use a single queue for pppoe traffic so only a single CPU core which limits throughput.
It doesn't effect em NICs because they are single queue anyway, equally limited performance for all protocols. igb is multi-queue for anything it can use RSS on which is IP traffic.
I'm not aware of any NIC that can hash pppoe traffic to use RSS.https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=203856#c11
Steve
-
Thanks for the reply Steve, then I may as well go with the newer NICs in case I change to an ISP that doesn't use PPPoE.
-
Or at least make sure you get a CPU with the best single thread rating you can. The i3 is significantly faster.
Steve
-
Rumours say the Intel 700-series cards can do RSS even over that bloody PPPoE WAN. The tricky part is I wasnt able to find the proof word-by-word confirmation in that bloody intel spec sheet of the 700-series NIC chip (its a light 1600+ page madness, if any of you suffer insomnia).
You can go figure yourself: https://ark.intel.com/content/www/us/en/ark/products/189534/intel-ethernet-controller-x710-at2.html?wapkw=X710-AT2Additionally, Microsoft (owner of the RSS algorithm specification) does not consider PPPoE as a valid Toeplitz hash-generating scenario either. You can go figure yourself:
https://docs.microsoft.com/en-us/windows-hardware/drivers/network/rss-hashing-typesSo either the whole industry is keeping these secrets, or everybody is just plain clueless in this topic since a decade. Until then, you can forget to have underpowered Atom or AMD G-series x86/x64 1Ghz-class CPUs (no matter if they are 4-8-or 1024 cores) perform well in routing of Gbit/sec traffic on single core.
-
@soder said in Gigabit PPPoE and Intel Drivers:
Rumours say the Intel 700-series cards can do RSS even over that bloody PPPoE WAN. The tricky part is I wasnt able to find the proof word-by-word confirmation in that bloody intel spec sheet of the 700-series NIC chip (its a light 1600+ page madness, if any of you suffer insomnia).
The 700 series has programmable hash functions so you can make them all sorts of things. Whether your OS will do something useful with it is a different question. The card will probably draw more power than a CPU that's challenged by this, so it's a moot point.
So either the whole industry is keeping these secrets, or everybody is just plain clueless in this topic since a decade. Until then, you can forget to have underpowered Atom or AMD G-series x86/x64 1Ghz-class CPUs (no matter if they are 4-8-or 1024 cores) perform well in routing of Gbit/sec traffic on single core.
Works fine on some OS's, so let's be clear that this is an implementation issue in freebsd.
-
@VAMike Do you have first-hand experience?
-
@soder said in Gigabit PPPoE and Intel Drivers:
@VAMike Do you have first-hand experience?
yes. you can drop a linux fw (for example) onto the same system and get higher pppoe performance because the pppoe implementation isn't single threaded. I haven't tested against the mpd5 implementation on freebsd, which I assume would be faster than the baseline.
-
@VAMike Of course if the CPU in said low performance router cannot keep up with the gbit PPPoE traffic, it is very unlikely that buying a 700-series card financially makes sense. So its only about theory..
-
@soder said in Gigabit PPPoE and Intel Drivers:
@VAMike Of course if the CPU in said low performance router cannot keep up with the gbit PPPoE traffic, it is very unlikely that buying a 700-series card financially makes sense. So its only about theory..
Either I'm misunderstanding you or you're misunderstanding me. I was not talking about dropping a 700-series card into a low powered device to improve PPPoE performance.