smlParallel.parmap_queue

STRUCTURE

SYNOPSIS

Run a function #function with parameter 'a in parallel on a list of arguments 'b list.

DESCRIPTION

The first argument of parmap_queue_extern precises the number of cores. To be efficient, each of worker is executed on its own process. Thus, I/O functions for parameter, argument list and result need to be specfied in the structure extspec. The communications files are stored under the directory #parallel_dir. As soon as a worker is available, a new job is started on the next argument of the list. Since the workers are executing from a script, it is possible to execute arbitrary side-effect code before running #function by specifing #reflect_globals. The directory where the signature containing extspec is located has to be given in #self_dir and its full sml identifier in #self (see Example).

FAILURE

Never fails but can produce uncaught exceptions causing the main thread to be stuck in a loop. See buildheap_ files in subdirectories of #parallel_dir for observing the standard out of #function.

EXAMPLE

(*
(* in foo.sml *)
val doublespec : (unit,int,int) extspec =
  {
  self_dir = change "directory where foo.sig is located",
  self = "foo.doublespec",
  parallel_dir = default_parallel_dir,
  reflect_globals = "()",
  function = let fun f _ (x:int) = 2 * x in f end,
  write_param = let fun f _ () = () in f end,
  read_param = let fun f _ = () in f end,
  write_arg = let fun f file arg = writel file [its arg] in f end,
  read_arg = let fun f file =
     string_to_int (singleton_of_list (readl file)) in f end,
  write_result = let fun f file r = writel file [its r] in f end,
  read_result = let fun f file = string_to_int (hd (readl_rm file)) in f end
  }

(* in foo.sig *)
val doublespec : (unit,int,int) extspec
*)

- load "smlParallel"; open smlParallel; load "foo"; open foo;
(* output omitted *)
> val it = (): unit

- parmap_queue_extern 2 doublespec () [1,2,3];
(* output omitted *)
> val it = [2, 4, 6]: int list

COMMENTS

If a user requires shared memory between workers, it is possible to rely on smlParallel.parmap_queue (or Parmap.parmap). However, in general it is quite difficult to obtain a significant speed-up when calling smlParallel.parmap_queue (or Parmap.parmap). One reason is that the threads need to be stopped temporarily during garbage collection.

SEEALSO

exec_script