DSP like TCP/IP processing in linux kernel
From: Indrek Kruusa
Date: Tue Jan 11 2005 - 20:19:47 EST
Hi!
My goal is to:
- learn more about DSP like TCP/IP processing
- investigate network/transport layer implementation in Linux
- investigate the possibility to increase network performance by SIMD
processing
- maybe compile some kind of document from my work
- ask lot of questions here :)
There are lot of quite basic questions to think about. The most basic is
how to cope with SIMD at all.
Hardware
----------
CPU + ext. == DSP?
(+) SIMD (vertical/horisontal vectorized processing)
(-) (??) h/w MAC/shifter/flow control, separate x/y path
* ARM/MIPS/PowerPC/SPARC/x86 - they all have (or planned) somekind of
SIMD extension but how much theory I can borrow from DSP world for SIMD
CPUs? (quite a lot I think)
* network processors (Intel,Broadcom etc.) - can I borrow something from
there? (only maybe how/why work is distributed between multiple cores)
Software (compiler)
--------------------
Is it possible at all to build a framework to cover those different
ext.'s in common way?
* gcc builtins for particular arch (well, arm,x86,powerpc are there)
* intrinsics. only intel (?)
* somekind of layer above gcc builtins? (ehh...)
Tools
-------
How should I produce SIMD code?
* Which tools I have to work with SIMD ext.? (umm...vim)
* Which way I should choose: "look, feel and think" or pure math and
modelling? (I don't like J.B.J. Fourier...)
Method
-------
How to start?
* tracing, profiling - squeeze out the performance hotspot and optimize
it (is the underlying design suitable at all for this?)
* sketch down current implementation in linux and think again about
whole thing (results after one year...for me :) )
* in networking context: start with the most common part: calculate
checksums for IP and TCP (I like this idea)
* Forget about it: turn on -ftree-vectorize when gcc 4.x is here (no-no...)
* Forget about it: this kind of tweaking is not for Linux kernel (what
about RAID-6 driver? :) )
Well, before I look into Documentation/, kernelnewbies.org, source etc.
I'd like to ask: is it worth to go further? At least I hope so :)
thanks,
Indrek
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/