On 10/8/07, Charles Forsyth <[EMAIL PROTECTED]> wrote:
> > It may be the case that we have cut our
> > abstractions too high to take advantage of some architectural features
> > present in modern microprocessors -- it may be that we want to allow
>
> I might draw a different conclusion. As with some peculiar memory subsystems,
> botched device interfaces, or 80 core processors, I'd say that perhaps the
> architectural features are there to meet the needs of hardware designers and
> are
> not actually designed to run the programs people are actually writing.
> Perhaps
> someone told them the software people wanted some of this stuff, but I've got
> my doubts!
>
I would be tempted to draw the same conclusion if other
implementations didn't seem to do a better job. Granted, our existing
points of comparison are highly specialized, but I still believe we
can accommodate both specialized and generalized cases as well as find
middle ground which may make the most sense for applications.
Further, I wonder if some of our locking interfaces may be better
built using some of the experiences of the past 20 years -- in
particular for larger SMP. Also, I wonder if we have the right set of
abstractions for our locking and queueing primitives for all
architectures. In particular -- if massive multi-core architectures
evolve to support IPC more natively, we'll want to take advantage of
those primitives.
Another examples is whether or not we could make better use of the
reservation based schemes on modern PPC versus the TAS model. It just
seems like given the architectural differences between the various
platforms we support, we might be able to do better native versions of
higher level primitives rather than just providing uniform support for
lower level primitives and always basing the higher level primitives
on that.
> Perhaps someone told them the software people wanted
> some of this stuff, but I've got my doubts!
Well, we can put ourselves in the position of requesting that stuff.
Designs for next generation Blue Genes and Power processors are still
in flux. One of the downsides of focusing on the available generation
is that we are losing out on the ability to try and affect the next
generation(s).
-eric
