On Thu, 9 Dec 1999, jcordina wrote:
> > > Well the facts talk for themselves. Linux 2.2.12 running on a single
> > > processor is able to context switch in around 1us. A user level thread
> > > system which is already built and tested and works perfectly can do a
> > > context switch in 55ns !!!
> >
> > Hrmm, this really makes me curious about
> > the size of the chunks of work you let
> > the threads do :)
> >
> What does the size of the thread work have to do with context switching when
> merely all the processor registers are saved on the context switch??
If you're doing work in 5ms portions, the context
switch will only be 0.5% overhead. I wouldn't
worry about it until you're hitting 2% or something.
> > If there is a free processor, it will be used. Linux knows
> > how much time it costs to move a thread from one CPU to an
> > other (cache flush time) and can make quite informed
> > decisions about when to move a process to where.
>
> Yes but once a thread is allocated to one processor it is a strategy
> that is should not migrate it to another processor exactly due to
> cache consistency. Then what happens if all threads on the process are
> on one processor ??
Won't happen. As I explained before Linux will
move the threads around as needed, avoiding
unneeded cache thrashing where possible.
> Also if the thread is interrupted by something else in the system then
> Linux knows best and will be allowed to schedule the interruptions as
> it sees fit. Yet I am not about rewriting the Linux scheduler I am on
> about scheduling my threads as I see fit
- you don't have to rewrite the scheduler, it already does what you
need (AFAIK)
- you cannot 'see fit' because only the kernel knows
enough to make informed scheduling decisions
regards,
Rik
--
The Internet is not a network of computers. It is a network
of people. That is its real strength.
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/dmentre/smp-howto/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
