• Chris Dew (25/25) Sep 28 2011 Warning: I've written multi-threaded code in C, Haskell and Java, but

Chris Dew <cmsdew gmail.com> writes:

Warning: I've written multi-threaded code in C, Haskell and Java, but
only single threaded code in D.

I'll use the word 'fibre' for a job which may take some time to run,
and may be blocked from time to time.  It may mean the same as 'task'
in D, but I'm not sure.  Greenlets, lightweightthreads or possibly
cooroutines may be other names which have some bearing here.

In order to run N fibers efficiently I'm making a big assumption that
we need to map them to a pool of M threads, where M is ~= NumCores (+
num blocked threads?) and then use some form of 'select' or 'epoll' in
the runtime to swap them in and out of threads, when they are really
to run again.

What I think I want is fibre local storage, by default.  Thread local
storage by default sounds fine for low-level work, but I don't want to
think about, or manually implement, how my fibres are mapped onto
threads efficiently.

As fibres can be moved between threads by the runtime, does that mean
that their storage must be 'shared', with all the inefficiency that
brings?

Are there Uniqueness Types in D?  They could be used to move the
fibre's storage between threads as the fibre moves, without
compromising safety.

If not, how do you currently implement storage for a fibre (task?), to
make it migrateable when the runtime moves the fibre between threads?

Thanks,

Chris.