blog.osau.re

MirageOS and OCaml stuffs.
Back to index

Cooperation and Lwt.pause

Here's a concrete example of the notion of availability and the scheduler used (in this case Lwt). As you may know, at Robur we have developed a unikernel: opam-mirror. It launches an HTTP service that can be used as an OPAM overlay available from a Git repository (with opam repository add <name> <url>).

The purpose of such an unikernel was to respond to a failure of the official repository which fortunately did not last long and to offer decentralisation of such a service. You can use https://opam.robur.coop!

It was also useful at the Mirage retreat, where we don't usually have a great internet connection. Caching packages for our OCaml users on the local network has benefited us in terms of our Internet bill by allowing the OCaml users to fetch opam packages over the local network instead of over the shared, metered 4G Internet conncetion.

Finally, it's a unikernel that I also use on my server for my software reproducibility service in order to have an overlay for my software like Bob.

In short, I advise you to use it, you can see its installation here (I think that in the context of a company, internally, it can be interesting to have such a unikernel available).

However, this unikernel had a long-standing problem. We were already talking about it at the Mirleft retreat, when we tried to get the repository from Git, we had a (fairly long) unavailability of our HTTP server. Basically, we had to wait ~10 min before the service offered by the unikernel was available.

Availability

If you follow my articles, as far as Miou is concerned, from the outset I talk of the notion of availability if we were to make yet another new scheduler for OCaml 5. We emphasised this notion because we had quite a few problems on this subject and Lwt.

In this case, the notion of availability requires the scheduler to be able to observe system events as often as possible. The problem is that Lwt doesn't really offer this approach.

Indeed, Lwt offers a way of observing system events (Lwt.pause) but does not do so systematically. The only time you really give the scheduler the opportunity to see whether you can read or write is when you want to... read or write...

More generally, it is said that Lwt's bind does not yield. In other words, you can chain any number of functions together (via the >>= operator), but from Lwt's point of view, there is no opportunity to see if an event has occurred. Lwt always tries to go as far down your chain as possible:

Lwt is rather sparse in adding cooperation points besides Lwt.pause and read/write operations, in contrast with Async where the bind operator is a cooperation point.

If there is no I/O, do not wrap in Lwt

It was (bad1) advice I was given. If a function doesn't do I/O, there's no point in putting it in Lwt. At first glance, however, the idea may be a good one. If you have a function that doesn't do I/O, whether it's in the Lwt monad or not won't make any difference to the way Lwt tries to execute it. Once again, Lwt should go as far as possible. So Lwt tries to solve both functions in the same way:

val merge : int array -> int array -> int array

let rec sort0 arr =
  if Array.length arr <= 1 then arr
  else
    let m = Array.length arr / 2 in
    let arr0 = sort0 (Array.sub arr 0 m) in
    let arr1 = sort0 (Array.sub arr m (Array.length arr - m)) in
    merge arr0 arr1

let rec sort1 arr =
  let open Lwt.Infix in
  if Array.length arr <= 1 then Lwt.return arr
  else
    let m = Array.length arr / 2 in
    Lwt.both
      (sort1 (Array.sub arr m (Array.length arr - m)))
      (sort1 (Array.sub arr 0 m))
    >|= fun (arr0, arr1) ->
    merge arr0 arr1

If we trace the execution of the two functions (for example, by displaying our arr each time), we see the same behaviour whether Lwt is used or not. However, what is interesting in the Lwt code is the use of both, which suggests that the processes are running at the same time.

"At the same time" does not necessarily suggest the use of several cores or "in parallel", but the possibility that the right-hand side may also have the opportunity to be executed even if the left-hand side has not finished. In other words, that the two processes can run concurrently.

But factually, this is not the case, because even if we had the possibility of a point of cooperation (with the >|= operator), Lwt tries to go as far as possible and decides to finish the left part before launching the right part:

$ ./a.out
sort0: [|3; 4; 2; 1; 7; 5; 8; 9; 0; 6|]
sort0: [|3; 4; 2; 1; 7|]
sort0: [|3; 4|]
sort0: [|2; 1; 7|]
sort0: [|1; 7|]
sort0: [|5; 8; 9; 0; 6|]
sort0: [|5; 8|]
sort0: [|9; 0; 6|]
sort0: [|0; 6|]

sort1: [|3; 4; 2; 1; 7; 5; 8; 9; 0; 6|]
sort1: [|3; 4; 2; 1; 7|]
sort1: [|3; 4|]
sort1: [|2; 1; 7|]
sort1: [|1; 7|]
sort1: [|5; 8; 9; 0; 6|]
sort1: [|5; 8|]
sort1: [|9; 0; 6|]
sort1: [|0; 6|]

1: However, if you are not interested in availability and would like the scheduler to try to resolve your promises as quickly as possible, this advice is clearly valid.

Performances

It should be noted, however, that Lwt has an impact. Even if the behaviour is the same, the Lwt layer is not free. A quick benchmark shows that there is an overhead:

let _ =
  let t0 = Unix.gettimeofday () in
  for i = 0 to 1000 do let _ = sort0 arr in () done;
  let t1 = Unix.gettimeofday () in
  Fmt.pr "sort0 %fs\n%!" (t1 -. t0)

let _ =
  let t0 = Unix.gettimeofday () in
  Lwt_main.run @@ begin
    let open Lwt.Infix in
    let rec go idx = if idx = 1000 then Lwt.return_unit
      else sort1 arr >>= fun _ -> go (succ idx) in
    go 0 end;
  let t1 = Unix.gettimeofday () in
  Fmt.pr "sort1 %fs\n%!" (t1 -. t0)

$ ./a.out
sort0 0.000264s
sort1 0.000676s

This is the fairly obvious argument for not using Lwt when there's no I/O. Then, if the Lwt monad is really needed, a simple Lwt.return at the very last instance is sufficient (or, better, the use of Lwt.map / >|=).

Cooperation and concrete example

So Lwt.both is the one to use when we want to run two processes "at the same time". For the example, ocaml-git attempts both to retrieve a repository and also to analyse it. This can be seen in this snippet of code.

In our example with ocaml-git, the problem "shouldn't" appear because, in this case, both the left and right side do I/O (the left side binds into a socket while the right side saves Git objects in your file system). So, in our tests with Git_unix, we were able to see that the analysis (right-hand side) was well executed and 'interleaved' with the reception of objects from the network.

Composability

However, if we go back to our initial problem, we were talking about our opam-mirror unikernel. As you might expect, there is no standalone MirageOS file system (and many of our unikernels don't need one). So, in the case of opam-mirror, we use the ocaml-git memory implementation: Git_mem.

Git_mem is different in that Git objects are simply stored in a Hashtbl. There is no cooperation point when it comes to obtaining Git objects from this Hashtbl. So let's return to our original advice:

don't wrap code in Lwt if it doesn't do I/O.

And, of course, Git_mem doesn't do I/O. It does, however, require the process to be able to work with Lwt. In this case, Git_mem wraps the results in Lwt as late as possible (as explained above, so as not to slow down our processes unnecessarily). The choice inevitably means that the right-hand side can no longer offer cooperation points. And this is where our problem begins: composition.

In fact, we had something like:

let clone socket git =
  Lwt.both (receive_pack socket) (analyse_pack git) >>= fun ((), ()) ->
  Lwt.return_unit

However, our analyse_pack function is an injection of a functor representing the Git backend. In other words, Git_unix or Git_mem:

module Make (Git : Git.S) = struct
  let clone socket git =
    Lwt.both (receive_pack socket) (Git.analyse_pack git) >>= fun ((), ()) ->
    Lwt.return_unit
end

Composability poses a problem here because even if Git_unix and Git_mem offer the same function (so both modules can be used), the fact remains that one will always offer a certain availability to other services (such as an HTTP service) while the other will offer a Lwt function which will try to go as far as possible quite to make other services unavailable.

Composing with one or the other therefore does not produce the same behavior.

Where to put Lwt.pause?

In this case, our analyse_pack does read/write on the Git store. As far as Git_mem is concerned, we said that these read/write accesses were just accesses to a Hashtbl.

Thanks to Hannes' help, it took us an afternoon to work out where we needed to add cooperation points in Git_mem so that analyse_pack could give another service such as HTTP the opportunity to work. Basically, this series of commits shows where we needed to add Lwt.pause.

However, this points to a number of problems:

  1. it is not necessarily true that on the basis of composability alone (by functor or by value), Lwt reacts in the same way
  2. Subtly, you have to dig into the code to find the right opportunities where to put, by hand, Lwt.pause.
  3. In the end, Lwt has no mechanisms for ensuring the availability of a service (this is something that must be taken into account by the implementer).

In-depth knowledge of Lwt

I haven't mentioned another problem we encountered with Armael when implementing multipart_form where the use of stream meant that Lwt didn't interleave the two processes and the use of a bounded stream was required. Again, even when it comes to I/O, Lwt always tries to go as far as possible in one of two branches of a Lwt.both.

This allows us to conclude that beyond the monad, Lwt has subtleties in its behaviour which may be different from another scheduler such as Async (hence the incompatibility between the two, which is not just of the 'a t type).

Digression on Miou

That's why we put so much emphasis on the notion of availability when it comes to Miou: to avoid repeating the mistakes of the past. The choices that can be made with regard to this notion in particular have a major impact, and can be unsatisfactory to the user in certain cases (for example, so-called pure calculations could take longer with Miou than with another scheduler).

In this sense, we have tried to constrain ourselves in the development of Miou through the use of Effect.Shallow which requires us to always re-attach our handler (our scheduler) as soon as an effect is produced, unlike Effect.Deep which can re-use the same handler for several effects. In other words, and as we've described here, an effect yields!

Conclusion

As far as opam-mirror is concerned, we now have an unikernel that is available even if it attempts to clone a Git repository and save Git objects in memory. At least, an HTTP service can co-exist with ocaml-git!

I hope we'll be able to use it at the next retreat, which I invite you to attend to talk more about Lwt, scheduler, Git and unikernels!