[Cerowrt-devel] Infoworld: Google thoughts about speeding up internet

[Jim Gettys]

On 01/24/2012 04:50 PM, Richard Brown wrote:
> Infoworld has an article that points to a Google Blog that has a few suggestions for speeding up TCP connections. 
>
> Infoworld: http://www.networkworld.com/news/2012/012412-google-looks-to-speed-up-255231.html
> Google Blog: http://googlecode.blogspot.com/2012/01/lets-make-tcp-faster.html#comment-form
>
> It seems that these suggestions (which include increasing initial congestion window to 10 packets, and decreasing initial timeout to one second) would make bufferbloat worse in the outbound direction. 
>
> I am somewhat relieved that there are a couple comments to this effect in the blog page above. I don't know if the author understands the bufferbloat problem fully, though.
>
> Rich Brown
> Hanover, NH
>
I'll be talking with several Googler's about this next week. I sent
mail, as their blogger site defeated me. Youchung Cheng got back to me
promptly a few minutes ago.

I already wrote an ID about #1; IW 10 I think is evil unless
correspondingly fewer connections are used to make the situation no
worse than it already is for transient bufferbloat.

I'm actually most worried about #2; reducing the open timeout in the
face of what netalyzr shows would be a real bummer I would guess.  

I have no problems with #3. 

#4 might or might not be ok, but I think should be studied more
carefully in light of bufferbloat and the fact that congestion avoidance
is already having severe indigestion given the size of the buffers. 
Most people have been simulating the Internet the way it should be,
rather than the way it actually is, unfortunately.
                    - jim