Proposal: Relax rules for 'pure'

[Robert Jacques]

On Mon, 01 Nov 2010 10:24:43 -0400, Bruno Medeiros  
<brunodomedeiros+spam at com.gmail> wrote:
> On 29/10/2010 02:32, Robert Jacques wrote:
[snip]
>> The programming language Cilk popularized the concept of parallelization
>> through many small tasks combined with a work stealing runtime. Futures
>> are essentially the same concept, but because futures were generally
>> implemented with OS-threads, a thread pool or fibers/coroutines, that
>> term is generally avoided. Like message passing, tasks are often
>> implemented in libraries with Intel's threading building blocks probably
>> being the most famous, though both Microsoft's Task Parallel Library and
>> Apple's Grand Central are gaining mind-share. David Simcha currently has
>> a task library in review for inclusion to phobos. Basically, the point
>> of tasks is to provide parallelization with extremely low overhead (on
>> average a Cilk spawn is less than 4 function calls). That way, instead
>> of having a few coarse grain threads which neither scale nor load
>> balance well, you're encouraged to use tasks everywhere and therefore
>> reap the benefits of a balanced N-way scalable system.
>
> Hum, I see what you mean know, but tasks only help with the *creation  
> overhead* of otherwise spawning lots of OS threads, they don't solve the  
> main problems I mentioned.
> First, it may be fine to spawn 100 tasks, but there is still the issue  
> of deciding how many OS threads the tasks will run in! Obviously, you  
> won't run them in just one OS thread, otherwise you won't get any  
> parallelism. Ideally for your program only, your program would have as  
> many OS threads as there are cores. But here there is still the same  
> issue of whether its ok for you program to use up all the cores in your  
> machine. The compiler doesn't know that. Could it be enough to have a  
> global compiler option to specify that? I don't think so: What if you  
> want some code of your program to use as much OS-threads as possible,  
> but not some other code?
> Second, and perhaps more importantly, the very same issue occurs in the  
> scope of your program alone. So, even if you use all OS threads, and  
> don't care about other programs, spawning 100 tasks for some loop might  
> take time away from other more important tasks of your program. The  
> compiler/task-scheduler/whatever would not automatically know what is  
> acceptable and what is not. (the only exception being if your program  
> was logically single-threaded)
>

Controlling the task runtime thread-pool size is trivial. Indeed, you'll  
often want to reduce the number of daemon threads by the number of active  
program threads. And if you need fine grain control over pool sizes, you  
can always create separate pools and assign tasks to them. I think a  
reasonable default would be (# of cores - 2) daemons with automatic  
decreases/increases with every spawn/termination. But, all those settings  
should be controllable at runtime.