I know that in OS/2 Warp each thread is scheduled independently, which
means that in an SMP you'll have to assume that more than one of your
threads might be executing at the same time.
Agreed, if you write it with that assumption, it will be portable. (At
least from that point of view. :-)
Regards
Gustav
Dave Ihnat wrote:
>
> Gustav Schaffter wrote:
>
> > Interesting thoughts about context switching. But, correct me if I'm
> > wrong, I guess that in a multi processor environment, this may be quite
> > complicated since two (or more) threads are *really* executed in
> > parallel. Right?
>
> Well, maybe, but you can't know about that. I remember on the old Tandem
> multiprocessor boxes that we had to actually tell non-stop processes which
> processor they were to use, checkpoint events, etc.
>
> Threads are different. I don't know--and I'm dead certain that there are
> folk who will cheerfully and quickly inform me--if there are any thread
> libraries that let (or force) you to worry about multi-processor thread
> assignment. I do know that in many implementations the thread manages its
> own prioritization and execution *within that of the process*. That is,
> you may have multiple threads, but nothing can acquire clock ticks beyond
> those of the process; each thread shares with other threads within the
> execution allocation of the parent process.
>
> I don't know if anyone has done anything cool with SMP, in Linux or any
> other Unix, to allow a library routine (or, by extension, any function)
> in user space to indicate either what processor it wants to use, or
> that it can be assigned to a separate processor as a thread of
> execution. I don't see it in the online docs, but I'm not in a
> position right now to do a web search.
>
> In the normal course of things, in any case, you'll program your
> threads with respect to their own critical regions and dependencies.
> If the library can, it'll do assignment to individual processors with
> aid of the OS; if not, then so. You've got portable code, and
> depending on the services offered by the OS, you will or won't get the
> efficiencies of thread distribution across processors.
>
> Cheers,
> --
> Dave Ihnat
> [EMAIL PROTECTED]
--
pgp = Pretty Good Privacy. To get my public pgp key, send an e-mail to:
[EMAIL PROTECTED]
Visit my web site at http://www.schaffter.com
--
To unsubscribe: mail [EMAIL PROTECTED] with "unsubscribe"
as the Subject.
