Today's trick : memory limits with Gc alarms

Hello, today’s little known OCaml feature we will learn about are memory limits using Gc alarms.

Imagine you want to perform a memory-heavy computation, whose size of data set is not known in advance. (This post is inspired by the implementation of the SAT/SMT solver mSAT.) You would like this computation to fail gracefully if memory becomes insufficient. How can you do it?

The Out_of_memory exception

The first thing you might think about is the Out_of_memory exception. You simply run the computation inside an exception handler, expecting to receive this exception if memory ever becomes exhausted. However, for several reasons this is not reliable. First, the program is not always notified of failing allocations and gets killed instead much after, when it tries to access the memory for the first time (this behaviour of the operating system is called overcommitting). But this is the least problem, because you quickly figure out how to get rid of overcommitting, for instance by setting a hard memory limit to your process using ulimit.

However, you then encounter the second issue: in OCaml, if a minor allocation fails, then the program will issue a fatal error instead of raising Out_of_memory. That’s it: Out_of_memory is only reliable for allocations that are guaranteed to go directly on the major heap (by default those bigger than 256 words, or those performed for bigarrays or unmarshalling). Not for repeated small allocations on the minor heap, which would represent most allocations in a functional program. (There are bug reports about this, but making Out_of_memory reliable is reported to be non-trivial.)

GC alarms

GC alarms are callbacks which run at the end of major GC cycles. They are registered with Gc.create_alarm. Now observe this cool trick:

let run_with_memory_limit limit f =
  let limit_memory () =
    let mem = Gc.(quick_stat ()).heap_words in
    if mem > limit / (Sys.word_size / 8) then raise Out_of_memory
  let alarm = Gc.create_alarm limit_memory in
  Fun.protect f ~finally:(fun () -> Gc.delete_alarm alarm ; Gc.compact ())

That’s it, we simply raise Out_of_memory from the Gc alarm. You can try it in the toplevel on a memory-hungry function:

let f () =
  let x = ref [] in
  let rec f () = x := ()::!x ; f () in
  f ();;
run_with_memory_limit 1000000000 f;;


This is not reliable with multi-threaded programs, because we do not know in which thread the exception must be raised. This is fine for programs like mSAT and many logic/language tools, because such tools are rarely limited by blocking IO, so they do not lose much by remaining single-threaded in the absence of true parallelism. (Haskell proposes per-thread allocation limits as a built-in feature, which manifests itself likewise as an asynchronous exception.)


Behind each Gc alarm, there is a finalised block. The finaliser calls the callback, and sets itself to run again at the next cycle. Thus, the memory limit (thus implemented) is an example of asynchronous exception usefully arising from a finaliser. Was such a use of Gc alarms expected when they were conceived? History did not tell.


As the author of mSAT, I’d just like to say the trick of using gc alarms to implement a memory limit in mSAT was actually taken from the theorem prover zenon, where if I remember correctly it was written by @damiendoligez , which might explain the subtle use of alarms/finalisers ^^

Also, as a side note, for people who would like to implement a similar limit but for computation time, I’d advise against using alarms (as is done in the code of zenon iirc) as they have proven to be somewhat not reliable, and I’ve found using Unix timers to raise an Out_of_time exception much better:

let run_with_time_limit limit f =
  (* create a Unix timer timer *)
  let _ = Unix.setitimer Unix.ITIMER_REAL Unix.{it_value = limit; it_interval = 0.01 } in
  (* The Unix.timer works by sending a Sys.sigalrm, so in order to use it,
     we catch it and raise the Out_of_time exception. *)
  let () =
    Sys.set_signal Sys.sigalrm (
      Sys.Signal_handle (fun _ ->
          raise Out_of_time)
      ) in
  Fun.protect f ~finally:(fun () ->
    Unix.setitimer Unix.ITIMER_REAL Unix.{it_value = 0.; it_interval = 0. }

Thanks, you solved two mysteries: whether finalisers were originally intended to raise, and why GC alarms ended up in Pervasives (there are no examples in the docs, no use in the compiler, and the git logs give no explanation and refer to no issue on a bug tracker).

Zenon indeed made a blip on my grep radar, but I erroneously dismissed it as another program that only use alarms to print or log something. It shows me how hard it is to read a function and try to determine whether it raises or not.