dynamic TCP algorithms switching

[yan cui]

Hi all:

Currently, Linux has kinds of TCP congestion algorithms, such as
reno, cubic, bic, hybla, ...., and each TCP congestion algorithm has
its target networking environment. I just wonder to know is it
possible to do dynamic TCP algorithm switching? In other words, the
system has a combined TCP congestion algorithm (say, TCP-auto), and it
behaves like one of the integrated TCP congestion algorithms according
to different detected networking environment, but can switch to a
different one. For example, TCP-auto totally uses the set of
congestion control operations in TCP-cubic by default, but when it
detects that the current OS uses wireless networking, it switches to
some wireless friendly TCP congestion algorithm. Does Linux have some
features like that, or do you (networking developers and users) care
about it?

Best Wishes!

[Stephen Hemminger]

On Fri, 22 Nov 2013 17:49:58 -0500

Post by yan cui
Currently, Linux has kinds of TCP congestion algorithms, such as
reno, cubic, bic, hybla, ...., and each TCP congestion algorithm has
its target networking environment. I just wonder to know is it
possible to do dynamic TCP algorithm switching? In other words, the
system has a combined TCP congestion algorithm (say, TCP-auto), and it
behaves like one of the integrated TCP congestion algorithms according
to different detected networking environment, but can switch to a
different one. For example, TCP-auto totally uses the set of
congestion control operations in TCP-cubic by default, but when it
detects that the current OS uses wireless networking, it switches to
some wireless friendly TCP congestion algorithm. Does Linux have some
features like that, or do you (networking developers and users) care
about it?
Best Wishes!

You overestimate the advantage of one verus the other.
It is possible to control algorithm on per-socket, and per-route
but other than benchmarking there or bulk transfer for normal net
traffic Cubic works fine for all environments.

[yan cui]

Then, why include so many (current Linux has 10+ TCP congestion algorithms)
algorithms? For users who want to deploy their application on Linux
and if the applications
are system resource intensive, they always want to tune the
configurations of the operating systems for the last piece of
performance. If they do so, maybe they are confused
about which TCP congestion algorithm to use for their environment. So,
the only way is to try each algorithm one by one. I understand the
setting of the default TCP congestion
algorithm to be Cubic means that it works well for most environments.
But if others
are seldom used, or can be replace with another implementation.
Why not just remove from the kernel?

Yan

Post by Stephen Hemminger
On Fri, 22 Nov 2013 17:49:58 -0500

Post by yan cui
Currently, Linux has kinds of TCP congestion algorithms, such as
reno, cubic, bic, hybla, ...., and each TCP congestion algorithm has
its target networking environment. I just wonder to know is it
possible to do dynamic TCP algorithm switching? In other words, the
system has a combined TCP congestion algorithm (say, TCP-auto), and it
behaves like one of the integrated TCP congestion algorithms according
to different detected networking environment, but can switch to a
different one. For example, TCP-auto totally uses the set of
congestion control operations in TCP-cubic by default, but when it
detects that the current OS uses wireless networking, it switches to
some wireless friendly TCP congestion algorithm. Does Linux have some
features like that, or do you (networking developers and users) care
about it?
Best Wishes!

You overestimate the advantage of one verus the other.
It is possible to control algorithm on per-socket, and per-route
but other than benchmarking there or bulk transfer for normal net
traffic Cubic works fine for all environments.

[Stephen Hemminger]

On Fri, 22 Nov 2013 18:21:12 -0500

Post by yan cui
Then, why include so many (current Linux has 10+ TCP congestion algorithms)
algorithms? For users who want to deploy their application on Linux
and if the applications
are system resource intensive, they always want to tune the
configurations of the operating systems for the last piece of
performance. If they do so, maybe they are confused
about which TCP congestion algorithm to use for their environment. So,
the only way is to try each algorithm one by one. I understand the
setting of the default TCP congestion
algorithm to be Cubic means that it works well for most environments.
But if others
are seldom used, or can be replace with another implementation.
Why not just remove from the kernel?
Yan

Most are intended for research and testing only.
Only a few are worth considering in a production environment.
That is also why there so many qdisc algorithms as well.