Re: Processors in memory... NOT

News sbc <[EMAIL PROTECTED]> wrote:
>Most computation involves more than one data item.
>I think this implies that the optimum place to perform
>a computation will be at something resembling the
>centroid of the data items involved - weighted by
>access frequency.  With some allowance for treating
>instructions as data. During the life of a comnputation
>this "computational centroid" would shift over time.
>
>Parallel computations would be performed at something
>resembling a Viterbi constallation diagram.

The allowance for treating instructions as data may
be unnecessary; assuming that programs aren't self-modifying,
the typical mode of large parallel processing seems to be
huge datasets being run with code that is sized on the
of order of magnitude of the L2/L3 E$ available now.

Distributing local copies of the instructions to the
sea of processors is probably cheap and not too difficult.

Even for something today like Blue Gene, 4mm^2 will get you
about a megabyte of EDRAM on CU-08 from IBM's foundrys.
Building using 150 mm^2  chips you should be able to fit
16 sets of (PPC440+big FPUs+1meg EDRAM) per chip.
Assuming that a goodly chunk of current working data
plus the code core fits within a megabyte, of course.
The problem sets would need finer grained analysis.

The point that you make about data and computation
locality within the dataset is a good one though.  
And feeds in to a variant of the issue you bring
up in another posting, of parallelism and hiding
or exposing system resources.  The layout of the
data sets within the system's physical memory
will hugely affect the behaviour of such computational
models... the decision of how to clump data,
and where to put specific parts of it, will be
a major optimization depending completely on
what the actual system topology and latencies
are for large parallel systems.

There are already people edging towards acknowledging
that from the OS and hardware side.  I don't know what
other manufacturers are up to, but Sun's been talking
about enhancements in future OSes to enable the system
to migrate RAM contents closer to where it's being used.


-george william herbert
[EMAIL PROTECTED]