Was Moore's law, now something else, parallelism

Bill Bogstad bogstad-e+AXbWqSrlAAvxtiuMwx3w at public.gmane.org
Tue Jul 13 12:22:16 EDT 2010


On Tue, Jul 13, 2010 at 10:28 AM, Edward Ned Harvey
<lopser-Z8efaSeK1ezqlBn2x/YWAg at public.gmane.org> wrote:
>> From: discuss-bounces-mNDKBlG2WHs at public.gmane.org [mailto:discuss-bounces-mNDKBlG2WHs at public.gmane.org] On
>> Behalf Of Mark Woodward

>> With
>> many multiple CPUs, there is actually REAL benefit that can be taken
>> from it. Also, old truisms may now becoming wrong. A user space process
>> for handling services should now be effectively more efficient (in
>> operation) than kernel based ones as long as resource access and
>> contention are managed well.
>
> Now this one, I am completely missing.  How on earth can anything in user
> space be *more* efficient than in kernel space?
>
> I could see it, if you were saying "just as efficient" ... but "more"
> efficient?

If you assume lots of cores, then single-threaded monolithic kernels
will have lower throughput for some types of load then other designs
which can take advantage of the multiple cores.   Mark mentions some
of the possibilities.  Here's an (incomplete?) list:

1. Kernel threads
2. User space processes
3. Multi process/modular kernels

I'm not sure calling such designs more efficient is correct, but they
may very well improve throughput often enough that they become the
more dominant design paradigm.

Linux already uses #1 and #2 and it's not clear to me that most
benefits of multi-core can't be gotten without resorting to #3.

Bill Bogstad






More information about the Discuss mailing list