Skip to content

lucasaiu/ocaml

 
 

Repository files navigation

Reentrant OCaml

This branch is the first step towards multicore support in OCaml. The reentrant runtime allows multiple runtimes to run independantly in the same process, typically one per core on multicore machines, enabling parallel computation. Each runtime can have many threads associated, but threads do not run in parallel on each single runtime. This enables us to keep the current OCaml garbage collector, which is extremely efficient in the sequential case, essentially as it is.

Different runtimes can exchange messages using send and receive primitives on shared mailboxes.

The feature is also useful when multiple runtimes are exposed as C libraries.

Status

DONE

  • Most static variables of the runtime have been moved into the context
  • Most C functions of the runtime have been renamed to _r(CAML_R, to take as first parameter the runtime context containing all former static variables.
  • External functions in OCaml can now be declared as "reentrant", i.e. receiving the context as first parameter.
  • The AMD64 backend of ocamlopt has been updated to use the runtime context and provide it to reentrant external functions.
  • The bytecode interpreter now also uses the context data structure, and can be used for parallel programming.
  • C functions in unix, str, bigarray have been updated, static variables have been moved into library specific contexts.
  • Globals generated by ocamlopt are now local to each runtime
  • Context splitting and communication primitives work
  • Experimental high-level parallel programming support, with algorithmic skeletons
  • Fixed otherlibs/systhreads, to allow each runtime to have many (non-parallel) threads associated.

IN PROGRESS

  • Fixing otherlibs/threads, to add multi-threading-over-multi-context support to bytecode

TODO

  • Only the AMD64 backend has been implemented. Implement the other ones.
  • Only TLS (Thread Local Storage) is currenty supported.
  • At the end, we need to provide all the functions that were renamed in _r in their older format, for backward compatibility.
  • Update the rest of C functions in otherlibs/

Principles

Each function in the runtime takes as first argument a "runtime context", containing all the variables that used to be static. To decrease the size of the patch, the file context.h defines for each variable a macro that lookup the variable within the context. All these functions have a suffix _r, and the macro CAML_R defines their first argument.

Once all functions will have been rewritten like that, we should implement the former version (without the first argument) by looking up the context in the thread local storage, using the function "caml_get_thread_local_context()".

Since we needed a global lock for all the runtimes, it has been implemented as "caml_enter_blocking_section()" (which should still be used for exclusion within a given runtime), by introducing two functions "caml_enter_lock_section()" and "caml_leave_lock_section()". These functions don't do anything right now, but locking should be added within the thread libraries by redefining "caml_enter_lock_section_hook" and "caml_leave_lock_section_hook". [FIXME: these names are very counterintuitive. Why do we have to name "caml_enter_lock_section" and "caml_enter_blocking_section" so similarly? --Luca Saiu. The name "caml_enter_lock_section" was terrible to begin with: see http://d.hatena.ne.jp/camlspotter/20100309/1268111257 ]

For static variables within libraries, the library should define its own runtime context, and use "caml_get_library_context_r(...)" to access it. Such contexts are currently limited to 24 (MAX_OTHER_CONTEXTS), but the limitation should be removed soon.

About

OCaml "reentrant runtime" experimental branch

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • OCaml 82.0%
  • C 13.0%
  • Emacs Lisp 1.6%
  • Assembly 1.2%
  • Shell 0.6%
  • M 0.6%
  • Other 1.0%