Parmap is a minimalistic library allowing to exploit multicore architecture for
OCaml programs with minimal modifications: if you want to use your many cores to
accelerate an operation which happens to be a map, fold or map/fold
(map-reduce), just use Parmap's parmap
, parfold
and parmapfold
primitives in
place of the standard List.map
and friends, and specify the number of
subprocesses to use by the optional parameter ~ncores
.
See the example
directory for a couple of running programs.
Parmap is not meant to be a replacement for a full fledged implementation of parallelism skeletons (map, reduce, pipe, and the many others described in the scientific literature since the end of the 1980's, much earlier than the specific implementation by Google engineers that popularised them). It is meant, instead, to allow you to quickly leverage the idle processing power of your extra cores, when handling some heavy computational load.
The principle of parmap is very simple: when you call one of the three available primitives, map, fold, and mapfold , your OCaml sequential program forks in n subprocesses (you choose the n), and each subprocess performs the computation on the 1/n of the data, in chunks of a size you can choose, returning the results through a shared memory area to the parent process, that resumes execution once all the children have terminated, and the data has been recollected.
You need to run your program on a single multicore machine; repeat after me:
Parmap
is not meant to run on a cluster, see one of the many available
(re)implementations of the map-reduce schema for that.
By forking the parent process on a single machine, the children get access, for free, to all the data structures already built, even the imperative ones, and as far as your computation inside the map/fold does not produce side effects that need to be preserved, the final result will be the same as performing the sequential operation, the only difference is that you might get it faster.
The OCaml code is reasonably simple and only marginally relies on external C libraries: most of the magic is done by your operating system's fork and memory mapping mechanisms. One could gain some speed by implementing a marshal/unmarshal operation directly on bigarrays, but we did not do this yet.
Of course, if you happen to have open channels, or files, or other connections that should only be used by the parent process, your program may behave in a very wierd way: as an example, do not open a graphic window before calling a Parmap primitive, and do not use this library if your program is multi-threaded!
To obtain maximum speed, Parmap tries to pin the worker processes to a CPU, using the scheduler affinity interface that is available in recent Linux kernels. Similar functionality may be obtained on different platforms using slightly different API. Contributions are welcome to support those other APIs, just make sure that you use autoconf properly.
You can use Parmap in a native toplevel (it may be quite useful if you use the
native toplevel to perform fast interactive computations), but remember that you
need to load the .cmxs
modules in it; an example is given in example/topnat.ml
If the number of chunks is equal to the number of cores, it is easy to preserve
the order of the elements of the sequence passed to the map/fold operations, so
the result will be a list with the same order as if the sequential function would
be applied to the input. This is what the parmap
, parmapfold
and parfold
functions
do when the chunksize argument is not used.
If the user specifies a chunksize that is different from the number of cores,
the current implementation for parmap
, parmapi
, array_parmap
and
array_parmapi
by default does not guarantee the preservation of the order
of the results. If the keeporder
parameter is set to true, an alternative
implementation is used, that tags the chunks and reorders them at the end, so the result of
calling Parmap.parmap f l
is the same as List.map f l
. Depending on the
nature of your workload (in particular, number of chunks and size of the results),
this may be way more efficient than implementing a sorting mechanism yourself, but
may also end up using up to twice the space and time of the default implementation:
there is a tradeoff, and it is up to the user to choose the solution that better suits him/her.
No reordering logic is implemented for parmapfold
, parfold
and their
variants, as performing these operations in parallel only make sense if the
order is irrelevant.
In general, using little chunksize helps in balancing the load among the workers, and provides better speed, but incurs a little overhead for tagging and reordering the chunks: there is a tradeoff, and it is up to the user to choose the solution that better suits him/her.
Visiting an array is much faster than visiting a list, and conversion of an array to and from a list is expensive, on large data structures, so we provide a specialised version of map on arrays, that beaves exactly like parmap.
We also provide a highly optimised specialised parmap version that is targeted
to float arrays, array_float_parmap
, that allows you to perform parallel
computation on very large float arrays efficiently, without the boxing/unboxing
overhead introduced by the other primitives, including array_parmap
.
To understand the efficiency issues involved in the case of large arrays of float, here is a short summary of the steps that any implementation of a parallel map function must perform.
-
create a float array to hold the result of the computation. This operation is expensive: on an Intel i7, creating a 10M float array takes 50 milliseconds
ocamlnat Objective Caml version 3.12.0 - native toplevel # #load "unix.cmxs";; # let d = Unix.gettimeofday() in ignore(Array.create 10000000 0.); Unix.gettimeofday() -. d;; - : float = 0.0501301288604736328
-
create a shared memory area,
-
possibly copy the result array to the shared memory area,
-
perform the computation in the children writing the result in the shared memory area,
-
possibly copy the result back to the OCaml array.
All implementations need to do 1, 2 and 4; steps 3 and/or 5 may be omitted depending on what the user wants to do with the result.
The array_float_parmap
performs steps 1, 2, 4 and 5. It is possible to share steps
1 and 2 among subsequent calls to the parallel function by preallocating the result
array and the shared memory buffer, and passing them as optional parameters to the
array_float_parmap
function: this may save a significant amount of time if the
array is very large.
opam install parmap
make
make install
make test