Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Weaver]

And I have one and a half lines down currently. AA is perhaps mounting a massive assault on BTW and I’m just leaving them alone to get on with it. I have asked AA to see if they can put a stop to the madness with my constantly failing line #2. Line #4 is a bit sickly now too, with about 5.9 dB of HCD and a few ES downstream showing.

[Chrysalis]

Hub5 plugged in on VM. (ntl wall socket so old I had to not push cable too hard it felt was falling to bits)

connected to laptop

5.45pm Saturday

864 down 52 up. (ookla)
663 down single threaded, 910 multi (tbb)

So need to see how this goes for the 14 day grace and decide where I am going.  But this is a good start.

[Jon21]

Hub5 plugged in on VM. (ntl wall socket so old I had to not push cable too hard it felt was falling to bits)

connected to laptop

5.45pm Saturday

864 down 52 up. (ookla)
663 down single threaded, 910 multi (tbb)

So need to see how this goes for the 14 day grace and decide where I am going.  But this is a good start.

Nice speeds there. Glad you got a Hub 5, decent improvement over the Hub 4 even if it still feels a bit like the software on it is still in development. It's not fully displaying all the stats at the moment. No Docsis 3.1 channels displayed for example, although it shows that it's running in 3.1 mode.

[Chrysalis]

Thanks Jon21, In modem mode on PC tbb is 851 single threaded now with 929 multi.
770 from hetzner single threaded http
my hetzner server only 650 on cubic and about 620 on bbr.

I downloaded a patch on my xbox earlier,was a little under 300mbps.

[Alex Atkin UK]

Both Xbox and Playstation have crap download speeds in my experience.  Its pretty pathetic when you consider Steam Deck can max out my WiFi and at best clocks the CPU cores around 1Ghz slower than the consoles (as the CPU really gets hammered when downloading/decompressing/patching).

[Chrysalis]

Its a shame it isnt higher as really games is by far the biggest need for the speed.  But still almost 10x faster than what I was downloading at before (was throttled to 30 to prevent the VDSL from been swamped).  500mbyte patch downloaded in just over 10 seconds.

Since you was interested in bbr in my other thread, on FreeBSD the closest to bbr on bsd I think is cdg, you have to load the kernel module first before you can set it on sysctl (relevant for FreeBSD, opnsense and pfsense).

kldload cc_cdg
sysctl net.inet.tcp.cc.algorithm=cdg

[Alex Atkin UK]

From what I was reading though, TCP congestion control will do nothing on a router, as its only controlling how data is sent over a TCP connection established from THAT machine.  So about the only use on pfSense would be if you were using OpenVPN over TCP.

In my own testing, it doesn't impact downstream traffic, it improves a servers ability to quickly reach and maintain optimal transmission speed.  This is why it was such a great thing for Google as it optimised their servers performance.

BBR was relevant for my LAN because I have a NAS, so it improves the speed pulling data from that NAS (by using BBR on the NAS), or sending data to it (if BBR is used on the client).

[XGS_Is_On]

The only TCP stacks involved in congestion control, etc, are the client and the server. No router should be modifying anything apart from IP addresses and maybe source port number when it does NAT.

Need a TCP proxy for it to start interfering with or caring about congestion avoidance on a flow. Only time I've seen this done is by traffic shaping equipment to more gracefully control TCP throughput. Those and WAN optimisers that proxy TCP and use their own TCP stack across WAN and to each end.

[Chrysalis]

From what I was reading though, TCP congestion control will do nothing on a router, as its only controlling how data is sent over a TCP connection established from THAT machine.  So about the only use on pfSense would be if you were using OpenVPN over TCP.

In my own testing, it doesn't impact downstream traffic, it improves a servers ability to quickly reach and maintain optimal transmission speed.  This is why it was such a great thing for Google as it optimised their servers performance.

BBR was relevant for my LAN because I have a NAS, so it improves the speed pulling data from that NAS (by using BBR on the NAS), or sending data to it (if BBR is used on the client).

True yeah unless the router itself is doing stuff, i was just sharing the info in case you was interested, if its running its own squid or something thats when it would be more important.

[XGS_Is_On]

BBR was relevant for my LAN because I have a NAS, so it improves the speed pulling data from that NAS (by using BBR on the NAS), or sending data to it (if BBR is used on the client).

You really shouldn't need an exotic TCP stack across a sub-1 millisecond round trip lossless link unless your NAS is more like an HPC SAN. Should be zero loss and even when the link saturates selective acknowledgement should take care of it across that kind of latency with virtually zero throughput loss. Across a LAN with a standard networking stack 100 Gbit+ should be doable. I'm saturating a 25 Gbit/s DAC with no tweaking at all pulling data from an NVME cache to an NVME'd up server.

If you're actually seeing benefit across 10G you must have some pretty high resolution stats. What're you using?

[bogof]

770 from hetzner single threaded http

Sounds like you've had a bit of a result via Virgin.  Nice one.
Is that for Hetzner's own test files?  And is it stable for the duration of the 10GB file?
I've been having a look at their test files via AAISP with my new connection - it's up and down all over the place.
Probably averages to about 350-400, but there is a heck of a variation.  Better than Zen was via GEA, but that connection was poor in my estimation.
Graph from their own router, though it behaves the same if I use my own router.

[Alex Atkin UK]

You really shouldn't need an exotic TCP stack across a sub-1 millisecond round trip lossless link unless your NAS is more like an HPC SAN. Should be zero loss and even when the link saturates selective acknowledgement should take care of it across that kind of latency with virtually zero throughput loss. Across a LAN with a standard networking stack 100 Gbit+ should be doable. I'm saturating a 25 Gbit/s DAC with no tweaking at all pulling data from an NVME cache to an NVME'd up server.

If you're actually seeing benefit across 10G you must have some pretty high resolution stats. What're you using?

I might not, I just was getting some weird results when I first upgraded to 10Gbit for the NAS with the slower clients.  I had to enable flow control in the switches or Gigabit clients were tanking to something like 300Mbit, that doesn't seem necessary any more (I turned it off again out of curiosity) so maybe it was a glitch in the Linux kernel that has since been ironed out?

That said, Windows clients were even more unpredictable, so I just did whatever tweaks I could find to try and push Windows into playing nice.

I did manage to get my Mythic Beasts VPS upgraded to a newer OS, just for some bizarre reason you have to manually install Rocky Linux as its not in their pre-configure OS list.  So I have BBR on there too now.  I was going to just use Fedora (as its easier to upgrade between versions) but with Virtualmin not supporting it, seemed a hassle.

[Chrysalis]

My earlier xbox download test was on wifi by the way I forgot it wasnt on ethernet .

[Chrysalis]

Sounds like you've had a bit of a result via Virgin.  Nice one.
Is that for Hetzner's own test files?  And is it stable for the duration of the 10GB file?
I've been having a look at their test files via AAISP with my new connection - it's up and down all over the place.
Probably averages to about 350-400, but there is a heck of a variation.  Better than Zen was via GEA, but that connection was poor in my estimation.
Graph from their own router, though it behaves the same if I use my own router.

Yeah it was sustained, I think there is moderate evening peak issues on the upstream, but downstream from what I can tell is no visible contention.

The upstream issues should get resolved when they get round to docsis 3.1 US upgrades on my segment.

I made an adjustment on 'notsent_lowat' on my servers to bring bbr speeds up to cubic speeds. 

[bogof]

Yeah it was sustained, I think there is moderate evening peak issues on the upstream, but downstream from what I can tell is no visible contention.

The upstream issues should get resolved when they get round to docsis 3.1 US upgrades on my segment.

I made an adjustment on 'notsent_lowat' on my servers to bring bbr speeds up to cubic speeds.

Thanks, I did ask AAISP to see what they think of it.  It looks like the speed ramps up, and then something decides to choke it back down.  Though I must admit it does seem churlish to be complaining at these levels... lol