You are not logged in.

#1 2017-04-21 06:22:47

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Anyone played with their congestion control algo's ?

What the title says basically. It's been possible to switch them for a long time. See that or via adding a line to /etc/sysctl.conf I believe.

Why would you want to ? Well mostly for the sheer dorking of it. Though technically other reasons emerge. After playing with bbr (really briefly), it does seem to get better speeds than the usual default cubic. You'll have to check your kernels .config file in /boot to see which are supported. 

Some are supposedly specifically designed for certain applications or network conditions. Ie a lossy wireless network, rather than a solid one. Could make sense switching in such cases to get better speeds.

So what say thee ? Anyone with such experience or wants to do some comparison testing and share findings on this topic would be great.

Vll! smile

Last edited by BLizgreat! (2017-04-21 06:27:46)

Offline

#2 2017-04-21 06:31:32

Head_on_a_Stick
Member
From: London
Registered: 2015-09-29
Posts: 8,759
Website

Re: Anyone played with their congestion control algo's ?

Do you have any benchmarks to illustrate the difference between the various algorithms?

The best way to speed up the Linux networking stack is to switch to FreeBSD  devil


“Et ignotas animum dimittit in artes.” — Ovid, Metamorphoses, VIII., 18.

Forum Rules   •   How to report a problem   •   Software that rocks

Offline

#3 2017-04-21 06:41:06

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

^ Ohhhhhh you, you, you, Bsd'ers! big_smile

Nope been awhile since I've revisited this topic. Haven't looked for any recent side-by-sides. There's a ton of ways to tune the gnu/nix network stack.

Though as yet haven't tried a bunch on the 2dork list. Also haven't tried Bsd but know you wouldn't like it w/o reason(s) Hoas. smile

This kinda thing does make sense me thinks. Since it's so easy to switch and no harm, no foul, if someone wants to rollback to whichever instead, shrugs.

Some real world testing is what I'm trying to encourage with this dorkish thread. Was hoping you'd test the waters a bit my friend. wink

Last edited by BLizgreat! (2017-04-21 06:43:18)

Offline

#4 2017-04-21 06:48:02

Head_on_a_Stick
Member
From: London
Registered: 2015-09-29
Posts: 8,759
Website

Re: Anyone played with their congestion control algo's ?

BLizgreat! wrote:

Was hoping you'd test the waters a bit

Could be tricky...

Puffy: ~ $ uname -a
OpenBSD Puffy.lan 6.1 GENERIC.MP#6 amd64

big_smile


“Et ignotas animum dimittit in artes.” — Ovid, Metamorphoses, VIII., 18.

Forum Rules   •   How to report a problem   •   Software that rocks

Offline

#5 2017-04-21 06:55:13

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Noooooooooooo!

You're exclusively Bsd now?! Why Nix gawds, whyyyyyyy!? Am net restricted for the time being. So guess if any of the other nixers here feel the urge and then the urge to share findings, would be cool.

Either way think things like this are interesting and worth babbling about. smile

Offline

#6 2017-04-21 07:09:10

Head_on_a_Stick
Member
From: London
Registered: 2015-09-29
Posts: 8,759
Website

Re: Anyone played with their congestion control algo's ?

BLizgreat! wrote:

You're exclusively Bsd now?

No, I have just installed Windows 10 as well  devil

https://forums.bunsenlabs.org/viewtopic … 403#p49403

But seriously, I still have Debian for the family laptop but I prefer OpenBSD myself — I only use other operating systems for gaming.


“Et ignotas animum dimittit in artes.” — Ovid, Metamorphoses, VIII., 18.

Forum Rules   •   How to report a problem   •   Software that rocks

Offline

#7 2017-04-21 07:41:26

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Whew, that was close. big_smile

Good to hear for real Hoas. Do ya have any opinion on this junk ? Since you're currently offline here. Guessing you may be engaged in algo messing activities.

Vll! smile

Offline

#8 2017-04-21 17:12:16

Sector11
The Tpyo Knig Mod
From: 77345 ¡#
Registered: 2015-08-20
Posts: 5,442

Re: Anyone played with their congestion control algo's ?

Head_on_a_Stick wrote:

Do you have any benchmarks to illustrate the difference between the various algorithms?

The best way to speed up the Linux networking stack is to switch to FreeBSD  devil

Hahahahahahaha

It's like; The best way to speed up the Windows networking stack is to ...  devil


BunsenLabs Forum Rules ---== I'm a Conky 1.9'er ==---
System:    Host: d67 Kernel: 4.9.0-9-amd64 x86_64 (64 bit gcc: 6.3.0)
Desktop: Openbox 3.6.1 Distro: Debian GNU/Linux 9 (stretch)

Offline

#9 2017-04-21 17:18:59

Sector11
The Tpyo Knig Mod
From: 77345 ¡#
Registered: 2015-08-20
Posts: 5,442

Re: Anyone played with their congestion control algo's ?

BLizgreat! wrote:

Whew, that was close. big_smile

Good to hear for real Hoas. Do ya have any opinion on this junk ? Since you're currently offline here. Guessing you may be engaged in algo messing activities.

Vll! smile

Like nicking your self shaving in the early 1900's .. or you wife sticks her head in the bathroom suddenly with a: "BOO!" - again early 1900's.
That would be a direct 'slice'.  sad


BunsenLabs Forum Rules ---== I'm a Conky 1.9'er ==---
System:    Host: d67 Kernel: 4.9.0-9-amd64 x86_64 (64 bit gcc: 6.3.0)
Desktop: Openbox 3.6.1 Distro: Debian GNU/Linux 9 (stretch)

Offline

#10 2017-04-21 19:41:36

twoion
ほやほや
Registered: 2015-08-10
Posts: 2,459

Re: Anyone played with their congestion control algo's ?

For advanced traffic control on network interfaces, use the tc tool (`man tc`) to change the qdisc(s) attached to your interface. It's a complex topic; short intro here: https://wiki.archlinux.org/index.php/Ad … ic_control. This moves beyond simple congestion control algorithms, it allows you to shape traffic however you like. This is way more important than congestion control when it comes to 'fairly' allocating bandwidth to applications in your typical home network, as the situation is more likely 'I can't use Skype well while I download this 50G file' rather than 'I'm saturating my entire link capacity so I need to tweak TCP congestion control in order to optimize client behaviour' (see: DCTCP).


What can you see / On the horizon?

Offline

#11 2017-04-21 22:35:14

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Appreciate the input. Like all things gnu/nix can certainly be complex. Mentioned have noted a butt load of things that can be applied to tune a Nix network stack.

Though atm just wanted to try to pin down this topic. Bbr vs cubic specifically. Easy enough to switch em and back again.

Things like conduct free online speed tests a couple places, download 10-50mb whatever-files and see if one gets better speeds. Try the sucker out for x-period and form an opinion etc.

If it does provide a gain, then easy enough for a nixer to employ x-algo on their system. Just going on the basis that an easy gain, is an easy gain. So why not ?

Will get around to dorking with it firther. When it's more convenient fellows. smile

Last edited by BLizgreat! (2017-04-21 22:39:05)

Offline

#12 2017-04-22 06:06:03

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Twoion you were ahead of me fellow nixer. smile Someone does need tc-fq w this sucker for pacing, not quite the easy drop in id been hoping for afterall.

Found a bit of info on config'ing which looks relevant. Have to do some reading and that Arch link too, thanks.


Vll! smile

Last edited by BLizgreat! (2017-04-22 06:23:48)

Offline

#13 2017-04-22 10:44:13

Sun For Miles
Member
Registered: 2017-04-12
Posts: 137

Re: Anyone played with their congestion control algo's ?

Although the last time I've had *nix pose as a network device was in 2005/2006, there are a lot of products that are based on Linux kernel on the market today. And in a perfect world QoS should be implemented on every server. big_smile I've worked with various vendors and had different requests regarding QoS implementations. For me the best way to learn was with packet flow diagrams. This is a bit more visual resource for guys like me.


Señor Chang, why do you teach Spanish?

Online

#14 2017-04-22 11:01:35

twoion
ほやほや
Registered: 2015-08-10
Posts: 2,459

Re: Anyone played with their congestion control algo's ?

Sun For Miles wrote:

Although the last time I've had *nix pose as a network device was in 2005/2006, there are a lot of products that are based on Linux kernel on the market today. And in a perfect world QoS should be implemented on every server. big_smile I've worked with various vendors and had different requests regarding QoS implementations. For me the best way to learn was with packet flow diagrams. This is a bit more visual resource for guys like me.

That article is a very good find! Thanks.


What can you see / On the horizon?

Offline

#15 2017-04-22 23:39:21

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Another thanks and hello SunFM's.

Made some headway on this topic, seems one of those subjects not overly well documented. As to setting up fq for bbr.

Easy enough to set fq as default qdisc, though gets hazy as to what to do from there. Some additional info for other nixers here.

Bbr is included in kernel 4.9 or later. Not sure if that includes the ones included in Debian, it should, would check the .config file in /boot. Should also apply to backported kernels.

Also found out the fq qdisc is included, though not enabled as default qdisc in Debian 8 or later. Easy enough to do with the tc command or adding a line for it in /etc/sysctl.conf.

So atm that's about where I'm at on this.


Have seen nixers reporting substantial benefits to setting up bbr + fq when it comes to getting a better experience out of youtube and Google properties. Though since Google has purportedly rolled out bbr server-side, bbr has supposedly improved things for people w or w/o bbr client-side.

Vll! smile

Last edited by BLizgreat! (2017-04-23 00:57:21)

Offline

#16 2017-04-24 05:25:26

Sun For Miles
Member
Registered: 2017-04-12
Posts: 137

Re: Anyone played with their congestion control algo's ?

Interesting stuff. If I understood correctly your goal is to document performance of different algorithms since proper papers are lacking? I could just just suggest one of the tools for testing the performance of the NIC - standard starting point would be IPerf. smile

Mentioning Google is interesting since they have very curious implementation on their caching servers. All their services on CDN machines (they call them "GGC" servers - Google Global Cache) share same physical machines, just operate on different software layers. Sadly, complete info is very hard to find for outsiders, but there are presentations where they shared info that they've achieved 2, 3 and more times the performance at large scale traffic compared to popular vendors in the field (mainly Cisco).

Getting better performance out of streaming services is hard to measure because I am pretty sure ISP's often have various QoS policies that affect burst traffic. You would probably have at least some numbers to report if you set up real time traffic monitoring during testing period and measure the behaviour in that way.


Señor Chang, why do you teach Spanish?

Online

#17 2017-04-25 06:25:10

BLizgreat!
Resident Babbler - vll!
Registered: 2015-10-03
Posts: 1,016

Re: Anyone played with their congestion control algo's ?

Yeah was curious if there'd be a speed gain for personal users as a result of this congestion control algo and fq as qdisc.

Really no solid findings as yet. Based on one resource I'd found, bbr + fq did generally improve goodput but the retransmit rate they cited was ridiculous too.

Would have to have some impact on cpu overhead. Perhaps cubic and fq_codel are fine as is. Still think it's interesting but shelving this junk atm.

Offline

#18 2017-04-26 15:53:22

o9000
tint2 developer
From: Network Neighborhood
Registered: 2015-10-24
Posts: 398
Website

Re: Anyone played with their congestion control algo's ?

TCP 101: congestion control is performed by the sender, not the receiver. For a typical web client, the performance benefit of changing the TCP congestion control algorithm is exactly ZERO.

It's like expecting your mail to arrive faster because you installed a bigger letter box.

Offline

Board footer

Powered by FluxBB