future linux tcp

[emihaly]

hi, whene will be aviable final kernel 2.5 in linux? Have this kernel tcp/ip improvents over 2.4?

I miss > 64 kb recv/send buffers FreeBSD have 256 kb max buffer. And also i miss reduce nagle time from 0.2 to 0.1 ms . Today pc are fast and 0.2 is bigg time loss.

[RobSeace]

I don't think anyone knows when the 2.5 dev kernel will become the 2.6
stable kernel... And, I'm not really up on any network changes in 2.5; you'd
be best to go look around places like kernel.org or kerneltrap.org or the
kernel section on lwn.net's weekly editions...

However, I'm pretty sure you can have send/receive buffers greater than
64K, even as far back as the 2.2 kernel... The max values are configurable
via "/proc/sys/net/core/rmem_max" and "wmem_max" (and, the default
values via "[rw]mem_default")... Just set those to a higher value, and it
should work...

As for Nagle, I almost always just turn it off, so I don't really care about
it... In general, latency is usually more important to me than throughput...
And, a well-designed app generally shouldn't be doing a lot of separate
small write()'s/send()'s in a row, anyway, so there usually isn't even a
need for Nagle, I'd say...

[Uzume]

Nagle can be useful for more than just many small socket writes back to back. Large socket writes can be broken into transmission units based on MTU or other "magic" OS algorithms and these ends of the data can be combined.

An evaluation of its usefulness is of course still up to the application, testing, and in the end if the design/implmentation (AKA programmer).

[emihaly]

yes i know, but i think, 0.1 s is lot of time for todays cpu-s and 0.2 nagle is from old slow pc-ies.