Re: What is Parallel DML?

"Billy Verreynne" <vslabs_at_onwe.co.za> wrote in message news:9jdv42$2ml$1_at_ctb-nnrp1.saix.net...
> "Howard J. Rogers" <howardjr_at_www.com> wrote in message
>
> > ...er, which rather implies that the number of CPUs has some
> > relevance for the matter of whether or not it is effective.
>
> Multiple CPU's increases capacity - which is what parallel processing
 needs.
> However, you still have capacity on a single CPU. Thus you can effectively
> use parallel processing on a single CPU machine.

That doesn't get around the fact that at any one moment, only a single process can be serviced by a single CPU, regardless of capacity. OK, the cycles chewed up by the 'system idle time process' can be put to more effective use, but only one of those slave process can actually have time on the CPU at any given moment. And *that's* the reason my query slowed down with a degree of parallelism of 2 -having to perform a context switch between two processes on the one CPU is expensive.

Threads are a different game again, since there is no context switch when switching between threads within the one process.

Your mileage may vary: mine never has.

HJR
>
> Saying that parallel procesing does not work on a single CPU, is like
 saying
> that multithreading one a single CPU does not work. Neither is true. What
 is
> true that you have less capacity on a single CPU box which has less
 capacity
> than a dual CPU box, which has less capacity than a quad CPU box, and so
> on..
>
> Many database platforms runs their CPU/CPUs at far less than even 50%
> capacity. Which means that there is wasted CPU time that can very
> effectively be put to use - and parallel processing is one of the better
> methods to increase performance and make proper use of CPU capacity.
>
> > I've actually had a query *slow down* by specifying a degree of
> > parallelism on a single CPU box. And that wasn't a fantasy.
>
> The reason for that is that the amount of parallelism you used exceeded
 CPU
> capacity. And that can be done without even having to resort to parallel
> query or DML processing.
>
> --
> Billy
>
>
>
Received on Mon Jul 23 2001 - 04:13:39 CDT