[plug] SMP vs memory, was: *athon

[Greg Mildenhall]

On Fri, 23 Jul 1999, Mike Holland wrote:
> On Thu, 22 Jul 1999, Trevor Phillips wrote:
> > > But doesnt timesharing do a similar thing? If Vmware is hogging the CPU,
> > > other processes will be given priority and run fairly normally.
> > It's timesharing, except you have double the MHz to share amongst processes.
> > So if you have ONE process thrashing, then it can thrash away on one CPU,
> Yes, and I'm saying the thrashing process shouldnt affect the others,
> _unless_it_is_swapping_.
Close but no cigar. You also have to consider cache-thrashing, since the
process moves from a CPU whose cache has what the program needs in it, to
another CPU that has been running a different task. Shifting all that data
back and forth between the caches has a very bad effect on performance
- not as bad as swapping, but you can't get rid of it by adding more RAM.

> (Or possibly poor scheduling.)
Different scheduling, not poor scheduling. :)
Linux SMP is not really tuned for a desktop.

> I'm talking about a typical desktop case, with one CPU-bound process
> at a time, and lots of other processes. In particular, you want the GUI to
> respond quickly. The scheduler should give these light tasks all the CPU
> they want, and the CPU-bound task gets all the spare CPU, say 98%, so a
>  2nd CPU wont really help.
It does help, because it reducing cache-thrashing, as long as the
scheduler doesn't bounce the CPU-hog back and forth between processors.

-Greg