Re: what's next for the linux kernel?
From: Luke Kenneth Casson Leighton
Date: Wed Oct 05 2005 - 07:08:34 EST
On Tue, Oct 04, 2005 at 05:37:59PM -0400, Chris wrote:
> You write:
> | personally i find that i like a bit of a run-up and/or advance notice
> | of major paradigm shifts.
> I maintain that one of the problems is that we simply do not know how
> to efficiently (in all senses of the word) program multi-processing
> systems and applications. Various people have various theories about how
> it should be done, but they are not well-proven.
> We do know that the current approach is difficult and error-prone, but
> terribly attractive because of an at least superficial simplicity. So
> far none of the alternatives have been convincing enough to supplant it.
> Efficient changing of paradigms requires a target paradigm. In this
> environment, a good target paradigm does not appear to exist yet and
> may not be clear for years. Starting engineering work seems a little
> bit premature in that situation.
your words paraphrase nicely the issues faced - for software engineers.
however, if you were to speak to a hardware engineer - an embedded
systems designer of ASICs - they would have a completely different
take on it, because they are dealing with parallelism all the time,
_and_ they even have some tools to help them do that.
what those ASIC designers lack is an off-the-shelf affordable
multiprocessor chip that makes their parallel algorithms run
fast enough: they _have_ to go for custom ASIC.
which is where the pricing of 90nm - $250k mask charges making up most
of the $2m NREs - is so key. even .13micron you're looking at $200k
NREs [non-recoverable expenditure].
and if you were to do a chip at .13micron you would be
looking at 120 watts for something running at 800mhz with
only 1 million gates.
a pentium 3 800mhz in other words, and that won't exactly get you very
what the software engineers you refer to above lack is the toolchain to
assist them in developing for anything other than uniprocessor targets.
such tools and techniques are being researched and developed (i sent
references to a coupl and also a google search criteria in another
earlier LKML email): however, yet again, it's chicken-and-egg.
until the chips start going multiprocessor, the tools are going to
remain in research labs. until the tools come out of research labs,
the chips are going to remain useless.
where it all goes a bit pearshaped with that chicken-and-egg vicious
cycle is if the bottom drops out of 65nm and 45nm processes, such that
_even_ the top uniprocessor mass-market chip manufacturers are forced
down a parallel processing line.
my point is: we're starting to see evidence of that happening
(small-scale, 2-cores, 2-hyperthreads, talk of 4-cores, etc.
even the X-Box 360 PPC 3x2)
_therefore_, i invite people who do linux kernel development
to think ahead - to take a _lead_ for once instead of waiting
for hardware to drop into their laps, at which point it is once
again too late, the hardware design decisions will have
already been made by someone else, and you will be treated
like second class citizens. again.
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/