Pre-emptive vs. Cooperative Multitasking
Fri, 21 Jul 1995 19:44:54 -0700 (PDT)
On Fri, 21 Jul 1995, Francois-Rene Rideau wrote:
> ">>" is Jecel,
> ">" is Billy.
Add a layer of inderection to that, of course.
> >> - Cooperative Multitasking: Mostly avoids the need to explicitly
> >> control resource sharing. Allows you to violate "invariants"
> >> to boost performance as long as things are restored to normal
> >> before the next task switching point.
> > Note that all that good stuff is destroyed by the fact that you must poll
> > for tasks at all other times.
> Polling costs almost nothing as compared to the heavy cost of complete
> state saving. Saving the complete state costs thousands of cycles. If well
> implemented, polling costs some ten cycles each time. If even better
> implemented, it costs nothing at all (see URL above).
I'll have to take a look-- it doesn't make any sense how calling any
function could not take time.
I do agree that state saving is too expensive-- the best multithreader I
know of is probably so fast simply because it runs on 8*86es, with not
much state to save. Nonetheless, it does use some cooperative
state-saving techniques-- but no cooperative multitasking.
> > Essentially, the only advantage I can see for cooperative MT is during
> > short critical loops, and why not just explicitly give those loops
> > control?
> Cooperation is just *required* for persistency, inter-thread and
> distributed GC (remember, TUNES is to provide all that, too). Of course,
> you could use "conservative" GC and persistency, but that sures costs much
> more than cooperative schemes. By writing cooperative code, you pay once and
> for all when you write the compiler; by promoting anarchy, you must be
> paranoid, and you never cease to pay the cost.
I think we're operating on mixed wavelengths, possibly. I'm not an
anarchist, in code or otherwise; I simply see tasking as the domain of
the superobject (known by some OSes as the kernel).
In fact, one possible scheme that I think would fit with TUNES is to have
each object depend upon its superior for computing time. That way you
don't need a huge central authority-- only big enough to apportion power
among the topmost objects. You could have some object types that
delivered power preemptively, others that depended upon cooperation
> Cooperation is to follow laws, to increase information. It does *not*
> oppose preemption. It opposes absence of cooperation, which is anarchy.
I support preemption. I do not oppose cooperation.
Does TUNES have a motto? (Or was that it in your .sig that I just
deleted?) If not, may I suggest "rem non spem, factum non dictum",
meaning (in its original context) "an action, not hope; facts not speech",
but better translated here as "an object rather than a procedure, a fact
rather than a word".
The first part will serve to show others what the OS is-- a fully OO
system. The second will keep reminding us what to do about it. Sorry, I
forget the source for this quote, but I'll look it up if you feel you can