It doesn't seem that unreasonable to make a starting point of the median number of packets that Google's customers can cope with at once, and then fail down the way for those that have problems.
I'm not sure I would call it cheating. When I think of cheating, I think of a restricted action that gives an unfair advantage over competitors, of which is incredibly difficult for competitors to duplicate. But implementing the same hack, in this case, is easy, fairly common, and it isn't stepping outside of any real boundaries so it's not like the playing field is uneven. It also works and works well to provide users (customers) with a better experience.
The problem is that if everyone implements a very large initial congestion window that just causes more network load on networks that can't handle that amount of data all at once.
So it's not "cheating" in terms of your competitors, but it's "cheating" in the sense that either you're basically jumping the queue for your TCP flows or everyone else does the same thing and then things are back to square one, if networks can't deal with it.
An argument that networks now can deal with much larger congestion windows can be made, and Google is making it in the IETF, of course.
This has been on HN before. I read the article and comments as well as used the information at work.
If I had time I would go dig it out but unfortunately I'm leaving for a midterm. Maybe someone else can find it; I think it had a good discussion with it.
I didn't realise it had been on HN already. I just saw it because Sam Saffron tweeted it. I guess I should search more thorougly (I just searched for "cheat" and checked the first page.)
Don't sweat it, I didn't see it the first time and it's a good read, glad it was reposted. I'm sure I'm not the only one.
Also, HN search for previous submissions is hit or miss. A more reliable way is to search ddg or google for 'site:news.ycombinator.com [url you want to submit]'.
Recent kernels already have initcwnd and initrwnd set to 10. Since 2.6.38 or so. If you're using recent kernels (any 3.x), manually setting those to 10 won't change anything.
Somewhere in the early 2.6.3x, I think around .32 or .33, had the two patches to allow setting initcwnd and initrwnd as you demonstrate.
Somewhere in 2.6.35 or 2.6.36 era, initrwnd was changed to 10. Later, around 2.6.38, initcwnd was changed to 10.
I've seen people claim that certain web accelerators use an initcwnd more like 20 to 40.
CentOS 6.2 (the most recent version of a very popular server linux) is using 2.6.32
You can change initcwnd on CentOS 6.2 but not initrwnd. And it defaults to 3
People can check their server kernel version with
uname -a
Servers that appear to be at 20 or especially 40 might just be setting the buffers to match the maximum client side window (which is what Microsoft appears to do on their servers).
For loonix only sysctl's I guess it would be a matter of taste; however, if you want to keep your muscle memory strong for operating systems lacking full /proc support, you might prefer the sysctl command.
Slightly off topic: the Blogspot mobile theme is killing me with their swipe left and right gestures to navigate posts. When I'm scrolling down the page on my iPhone, I accidentally activate it constantly.
Ben fails to mention that the equation he proposed is "violated" is expressly "optional" and in fact the standard allows that "a TCP MAY start with a larger initial window".
From the article:
> Being non-standards-compliant in a way that privileges their flows relative to others seems more than a little hypocritical from a company that's making such a fuss about network neutrality.
That's completely off the mark. NN was a reaction to the plan by ISPs to throttle large sites (e.g. YouTube) that their own customers were requesting unless they were paid protection money. It later became a political thing, with people mostly deciding to support (or oppose) it based on party lines.
It has nothing to do with tuning one's network, in spite of rhetoric to the contrary. It was just an attempt to stop rent-seeking by greedy ISPs that has morphed politically after attracting lobbyist attention.
the microsoft behavior of disabling SS completely seems downright immoral to me. Who cares if the internet melts as long as they can deliver their home page a few ms faster. And it's not even a very interesting homepage
34 comments
[ 2.3 ms ] story [ 77.0 ms ] threadSo it's not "cheating" in terms of your competitors, but it's "cheating" in the sense that either you're basically jumping the queue for your TCP flows or everyone else does the same thing and then things are back to square one, if networks can't deal with it.
An argument that networks now can deal with much larger congestion windows can be made, and Google is making it in the IETF, of course.
http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-02
Anyway, super easy to try on linux servers.
First double check your existing setting
then if you want to enhance to IW10 then first check the last line of should be something like then copy it to the middle of this if you are using a VERY current linux kernel, you can also add initrwnd like so: Note the ip route change is not permanent unless you add it to something like /etc/rc.localYou should be able to see the change before/after in wireshark if you look very very carefully at the milliseconds before the first real ack.
If I had time I would go dig it out but unfortunately I'm leaving for a midterm. Maybe someone else can find it; I think it had a good discussion with it.
Op Follow up http://blog.benstrong.com/2010/11/slow-start-follow-up.html
Also, HN search for previous submissions is hit or miss. A more reliable way is to search ddg or google for 'site:news.ycombinator.com [url you want to submit]'.
Somewhere in the early 2.6.3x, I think around .32 or .33, had the two patches to allow setting initcwnd and initrwnd as you demonstrate.
Somewhere in 2.6.35 or 2.6.36 era, initrwnd was changed to 10. Later, around 2.6.38, initcwnd was changed to 10.
I've seen people claim that certain web accelerators use an initcwnd more like 20 to 40.
You can change initcwnd on CentOS 6.2 but not initrwnd. And it defaults to 3
People can check their server kernel version with
Servers that appear to be at 20 or especially 40 might just be setting the buffers to match the maximum client side window (which is what Microsoft appears to do on their servers).echo 1 | sudo tee /proc/foo
This has the added bonus that you can replace "echo 1" with a complex script that may not be safe to run as root.
For me, I can just copy over a modified sysctl.conf to a server and have such changes applied (reboot may be needed).
It's frustrating how often a website will provide a client-specific rendering which is much worse than the default. (I'm looking at you, Wordpress).
"This increased initial window is optional: a TCP MAY start with a larger initial window."
And specifically this: "MAY start with a larger initial window". "larger" than what?
a.) larger than that proposed?
or
b.) larger than was standard before this proposal?
That's completely off the mark. NN was a reaction to the plan by ISPs to throttle large sites (e.g. YouTube) that their own customers were requesting unless they were paid protection money. It later became a political thing, with people mostly deciding to support (or oppose) it based on party lines.
It has nothing to do with tuning one's network, in spite of rhetoric to the contrary. It was just an attempt to stop rent-seeking by greedy ISPs that has morphed politically after attracting lobbyist attention.