Parallel programming in Multicore OCaml

This tutorial will help you get started with writing parallel programs in Multicore OCaml. All the code examples along with their corresponding dune file are available in the code/ directory. The tutorial is organised into the following sections:

• Introduction
  • Installation
  • Compatibility with existing code
• Domains
• Domainslib
  • Task pool
  • Parallel for
  • Async-Await
    • Fibonacci numbers in parallel
• Channels
  • Bounded Channels
  • Task Distribution using Channels
• Profiling your code
  • Perf
  • Eventlog

Introduction

Multicore OCaml is an extension of OCaml with native support for Shared Memory Parallelism through Domains and Concurrency through Algebraic effects. It is slowly, but steadily being merged to trunk OCaml. Domains-only multicore is expected to land first followed by Algebraic effects.

Concurrency is how we partition multiple computations such that they can run in overlapping time periods rather than strictly sequentially. Parallelism is the act of running multiple computations simultaneously, primarily by using multiple cores on a multicore machine. The multicore wiki has comprehensive notes on the design decisions and current status of concurrency and parallelism in Multicore OCaml.

The Multicore OCaml compiler comes with two variants of Garbage Collector, namely a concurrent minor collector (ConcMinor) and a stop-the-world parallel minor collector (ParMinor). Our experiments have shown us that ParMinor performs better than ConcMinor in majority of the cases. ParMinor also does not need any changes in the C API of the compiler, unlike ConcMinor which breaks the C API. So, the consensus is to go forward with ParMinor during up- streaming of the Domains-only Multicore. ConcMinor is at OCaml version 4.06.1 and ParMinor has been promoted to 4.10.0 and 4.11.0. More details on the GC design and evaluation are available in this ICFP 2020 paper.

The Multicore ecosystem also has the following libraries to complement the compiler.

• Domainslib: Data and control structures for parallel programming.
• Lockfree: Lock-free data structures (list, hash, bag and queue).
• Reagents: Composable lock- free concurrency library for expressing fine grained parallel programs on Multicore OCaml
• Kcas: Multi-word compare and swap library

This tutorial takes you through ways in which one can profitably write parallel programs in Multicore OCaml. A reader is assumed to be familiar with OCaml, if not, they are encouraged to read Real World OCaml. The effect handlers story is not touched upon here, for anyone interested, do check out this tutorial and examples.

Installation

Up-streaming of the multicore bits to trunk OCaml in progress, with some PRs already merged to trunk. One can start using Multicore OCaml with the help of multicore-opam. Installation instructions for Multicore OCaml 4.10.0 compiler and domainslib can be found here. Other available compiler variants are here.

It will also be useful to install utop on your Multicore switch. opam install utop should work out of the box.

Compatibility with existing code

Multicore OCaml is compatible with existing OCaml code. It has support for the C API along with some tricky parts of the language like ephemerons and finalisers. To maintain compatibility with ppx there is a no-effect-syntax compiler variant in multicore-opam, that removes some syntax extensions.

Domains

Domains are the basic unit of parallelism in Multicore OCaml.

let square n = n * n

let x = 5
let y = 10

let _ =
  let d = Domain.spawn (fun _ -> square x) in
  let sy = square y in
  let sx = Domain.join d in
  Printf.printf "x = %d, y = %d\n" sx sy

Domain.spawn creates a new process of execution that runs along with the current domain.

Domain.join d blocks until the domain d runs to completion. If the domain returns a result after its execution, Domain.join d also returns that value. If it raises an uncaught exception, that is thrown. When the parent domain terminates, all other domains also terminate. To ensure that a domain runs to completion, we have to join the domain.

Note that the square of x is computed in a new domain and that of y in the parent domain.

Let us create its corresponding dune file and run this code.

(executable
  (name square_domain)
  (modules square_domain))

Make sure to use a multicore switch to build this and all other subsequent examples we encounter in this tutorial.

To execute the code:

$ dune build square_domain.exe
$ ./_build/default/square_domain.exe
x = 25, y = 100

So, as expected the squares of x and y are 25 and 100.

Common error message

Some common errors one might encounter while compiling Multicore code are

Error: Unbound module Domain

Error: Unbound module Atomic

Error: Library "domainslib" not found.

These errors usually mean that the compiler switch used to compile the code is not a multicore switch. Using a multicore compiler variant should resolve them.

Domainslib

Domainslib is a parallel programming library for Multicore OCaml. It provides the following APIs which enable easy ways to parallelise OCaml code with few modifications to sequential code:

• Task: Work stealing task pool with async/await parallelism and parallel_{for, scan}.
• Channels: Multiple Producer Multiple Consumer channels which come in two flavours, bounded and unbounded.

Domainslib is effective in scaling the performance when parallelisable workloads are available.

Task pool

In the Domains section, we saw how to run programs on multiple cores by spawning new domains. Often times we will find ourselves spawning and joining new domains numerous times in the same program, if we were to use that approach for executing code in parallel. Creating new domains is an expensive operation which we should attempt to limit however much possible. Task pool lets us to execute all parallel workloads in the same set of domains which are spawned at the beginning of the program. Let us see how to get task pools working.

Note: run #require "domainslib" with the hash before this, if you are running this on utop.

# open Domainslib

# let pool = Task.setup_pool ~num_domains:3
val pool : Task.pool = <abstr>

We have created a new task pool with three new domains. The parent domain is also part of this pool, thus making it a pool of four domains. After the pool is setup, we can use this pool to execute all tasks we want to run in parallel. The setup_pool function requires us to specify the number of new domains to be spawned in the task pool. The ideal number of domains to initiate a task pool with is equal to the number of cores available. Since the parent domain also takes part in the pool, the num_domains parameter should be one less than the number of available cores.

Closing the task pool after execution of all tasks, though not strictly necessary, is highly recommended. This can be done as

# Task.teardown_pool pool

Now the pool is deactivated and no longer usable, so make sure to do this only after all tasks are done.

Parallel for

parallel_for is a powerful primitive in the Task API which can be used to parallelise computations that use for loops. It can scale well with very little change to the sequential code.

Let us consider the example of matrix multiplication.

First, let us write down the sequential version of a function which performs matrix multiplication of two matrices and returns the result.

let matrix_multiply a b =
  let i_n = Array.length a in
  let j_n = Array.length b.(0) in
  let k_n = Array.length b in
  let res = Array.make_matrix i_n j_n 0 in
  for i = 0 to i_n - 1 do
    for j = 0 to j_n - 1 do
      for k = 0 to k_n - 1 do
        res.(i).(j) <- res.(i).(j) + a.(i).(k) * b.(k).(j)
      done
    done
  done;
  res

To make this function run in parallel, one might be inclined to spawn a new domain for every iteration in the loop, which would look like:

  let domains = Array.init i_n (fun i ->
    Domain.spawn(fun _ ->
      for j = 0 to j_n - 1 do
        for k = 0 to k_n - 1 do
          res.(i).(j) <- res.(i).(j) + a.(i).(k) * b.(k).(j)
        done
      done)) in
   Array.iter Domain.join domains

This will be disastrous in terms of performance majorly due to the fact that spawning a new domain is an expensive operation. What instead task pool offers us is, a finite set of available domains, which can be used to run your computations in parallel.

Arrays are usually more efficient compared with lists in the context of Multicore OCaml. Although they are not generally favoured in functional programming, using arrays for the sake of efficiency is a reasonable trade-off.

A better way to parallelise matrix multiplication with the help of a parallel_for.

let parallel_matrix_multiply pool a b =
  let i_n = Array.length a in
  let j_n = Array.length b.(0) in
  let k_n = Array.length b in
  let res = Array.make_matrix i_n j_n 0 in

  Task.parallel_for pool ~start:0 ~finish:(i_n - 1) ~body:(fun i ->
    for j = 0 to j_n - 1 do
      for k = 0 to k_n - 1 do
        res.(i).(j) <- res.(i).(j) + a.(i).(k) * b.(k).(j)
      done
    done);
  res

We can observe quite a few differences between the parallel and sequential versions. The parallel version takes an additional parameter pool, it is because, the parallel_for executes the for loop on the domains present in that task pool. While it is possible to initialise a task pool inside the function itself, it is always better to have a single task pool used across the entire program. As mentioned earlier, this is to minimise the cost involved in spawning a new domain. It is also possible to create a global task pool and use it across, but for the sake of being able to reason better about your code, it is recommended to use it as a function parameter.

We shall examine the parameters of parallel_for. It takes in pool as discussed earlier, start and finish as the names suggset are the starting and ending values of the loop iterations, body contains the actual loop body to be executed.

Parallel for also has an optional parameter chunk_size. It determines the granularity of tasks when executing them on multiple domains. If no parameter is given for chunk size, a default chunk size is determined which performs well in most cases. Only if the default chunk size doesn't work well, it is recommended to experiment with different chunk sizes. The ideal chunk_size depends on a combination of factors:

• Nature of the loop: There are two things to consider pertaining to the loop while deciding on a chunk_size to use, the number of iterations in the loop and amount of time each iteration takes. If the amount of time taken by every iteration is roughly equal, then the chunk_size could be number of iterations divided by the number of cores. On the other hand, if the amount of time taken is different for every iteration, the chunks should be smaller. If the total number of iterations is a sizeable number, a chunk_size like 32 or 16 is safe to use, whearas if the number of iterations is low, like say 10, a chunk_size of 1 would perform best.

• Machine: Optimal chunk size varies across machines and it is recommended to experiment with a range of values to find out what works best on yours.

Speedup

Let us find how the parallel matrix multiplication scales on multiple cores.

Speedup

The speedup vs core is enumerated below for input matrices of size 1024x1024.

We have achieved a speedup of 16 with the help of a parallel_for. It is very much possible to achieve linear speedups when parallelisable workloads are available.

Note that the performance of parallel code heavily depends on the machine, some machine settings specific to Linux systems for obtaining optimal results are described here.

Properties and Caveats of parallel_for

Implicit Barrier

The parallel_for has an implicit barrier, meaning other tasks if any, waiting to be executed after in the same pool will start only after all chunks in the parallel_for are done. So, we need not worry about creating and inserting barriers explicitly between two parallel_fors or some other operation after a parallel_for. Consider this scenario: we have three matrices m1, m2 and m3. We want to compute (m1*m2) * m3 where * indicates matrix multiplication. For the sake of simplicity, let us assume all three are square matrices of the same size.

let parallel_matrix_multiply_3 pool m1 m2 m3 =
  let size = Array.length m1 in
  let t = Array.make_matrix size size 0 in (* stores m1*m2 *)
  let res = Array.make_matrix size size 0 in

  Task.parallel_for pool ~chunk_size:(size/num_domains) ~start:0 ~finish:(size - 1) ~body:(fun i ->
    for j = 0 to size - 1 do
      for k = 0 to size - 1 do
        t.(i).(j) <- t.(i).(j) + m1.(i).(k) * m2.(k).(j)
      done
    done);

  Task.parallel_for pool ~chunk_size:(size/num_domains) ~start:0 ~finish:(size - 1) ~body:(fun i ->
    for j = 0 to size - 1 do
      for k = 0 to size - 1 do
        res.(i).(j) <- res.(i).(j) + t.(i).(k) * m3.(k).(j)
      done
    done);

    res

In a hypothetical situation where parallel_for did not have an implicit barrier, in the example above, it is very likely that the computation of res would not be correct. Since, we already have an implicit barrier, we will get the right computation.

Order of execution

for i = start to finish do
  <body>
done

A sequential for loop, like the one above, runs its iterations in the exact same order, from start to finish. In case of parallel_for the order of execution is arbitrary and varies between two runs of the exact same code. If the iteration order is important for your code to work as desired, it is advisable to use parallel_for with some caution.

Dependencies within the loop

If there are any dependencies within the loop, such as current iteration depends on the result of a previous iteration, odds are very high that the correctness of the code no longer holds if parallel_for is used. Task API has a primitive parallel_scan which might come in handy in scenarios like this.

Async-Await

Parallel for lets easily parallelise iterative tasks. Async-Await offers more flexibility to execute tasks in parallel which is especially useful in recursive functions. We have earlier seen how to setup and tear down a task pool. The Task API also has the facility to run specific tasks on a task pool.

Fibonacci numbers in parallel

We are going to calculate fibonacci numbers in parallel. First let us write a sequential function to calculate fibonacci numbers. This is a naive fibonacci function without tail-recursion.

let rec fib n =
if n < 2 then 1
else fib (n - 1) + fib (n - 2)

Observe that the two operations in recursive casefib (n - 1) and fib (n - 2) do not have any mutual dependencies which makes it convenient for us to compute them in parallel. Essentially, we can calculate fib (n - 1) and fib (n - 2) in parallel and add the results to get the answer.

We can do this by spawning a new domain for performing calculation and joining it to obtain the result. We have to be careful here to not spawn more domains than number of cores available.

let rec fib_par n d =
  if d <= 1 then fib n
  else
    let a = fib_par (n-1) (d-1) in
    let b = Domain.spawn (fun _ -> fib_par (n-2) (d-1)) in
    a + Domain.join b

We can as well use task pools to execute tasks asynchronously, which is less tedious and scales better.

let rec fib_par pool n =
  if n <= 40 then fib n
  else
    let a = Task.async pool (fun _ -> fib_par pool (n-1)) in
    let b = Task.async pool (fun _ -> fib_par pool (n-2)) in
    Task.await pool a + Task.await pool b

We can note some differences from the sequential version of fibonacci.

• pool is an additional parameter for the same reasons in parallel_for.

• if n <= 40 then fib n -> when the input is less than 40, we are running the sequential fib function. When the input number is small enough, it is better off to perform the calculations sequentially. We have taken 40 as the threshold here, some experimentation would help you to find a good enough threshold, below which the computation can be done sequentially.

• Task.async and Task.await are used to run the tasks in parallel.

  • Task.async executes the task in the pool asynchronously and it returns a promise, a computation that is not yet complete. After the execution runs to completion, its result will be stored in the promise.

  • Task.await waits for the promise to complete its execution and once it is done, the result of the task is returned. In case the task raises an uncaught exception, await also raises the same exception.

Channels

Bounded Channels

Channels act as medium to communicate data between domains. They can be shared between multiple sending and receiving domains. Channels in Multicore OCaml come in two flavours:

• Bounded: buffered channels with a fixed size. A channel with buffer size 0 corresponds to a synchronised channel, buffer size 1 gives the MVar structure. Bounded channels can be created with any buffer size.

• Unbounded: unbounded channels have no limit on the number of objects they can hold, they are only constrained by memory availability.

open Domainslib

let c = Chan.make_bounded 0

let _ =
  let send = Domain.spawn(fun _ -> Chan.send c "hello") in
  let msg =  Chan.recv c in
  Domain.join send;
  Printf.printf "Message: %s\n" msg

In the above example, we have a bounded channel c of size 0. Any send to the channel is blocked until a corresponding recv is encounterd. So, if we remove the recv in this example, the program would be blocking indefinitely.

open Domainslib

let c = Chan.make_bounded 0

let _ =
  let send = Domain.spawn(fun _ -> Chan.send c "hello") in
  Domain.join send;

The above example would be essentially blocking indefinitely because the send does not have a corresponding receive. If we instead create a bounded channel with buffer size n, it can store up to [n] objects in the channel without a corresponding receive, exceeding which the sending would block. We can try it with the same example as above just by changing the buffer size to 1.

open Domainslib

let c = Chan.make_bounded 1

let _ =
  let send = Domain.spawn(fun _ -> Chan.send c "hello") in
  Domain.join send;

Now, the send does not block anymore.

If you do not want to block in send or recv, send_poll and recv_poll might come in handy. They return a boolean value, if the operation was successful we get a true, otherwise a false.

open Domainslib

let c = Chan.make_bounded 0

let _ =
  let send = Domain.spawn(fun _ ->
          let b = Chan.send_poll c "hello" in
          Printf.printf "%B\n" b) in
  Domain.join send;

Since the buffer size is 0 and the channel cannot hold any object, this program prints a false,

The same channel may be shared by multiple sending and receiving domains.

open Domainslib

let num_domains = try int_of_string Sys.argv.(1) with _ -> 4

let c = Chan.make_bounded num_domains

let send c =
  Printf.printf "Sending from: %d\n" (Domain.self () :> int);
  Chan.send c "howdy!"

let recv c =
  Printf.printf "Receiving at: %d\n" (Domain.self () :> int);
  Chan.recv c |> ignore

let _ =
  let senders = Array.init num_domains
                  (fun _ -> Domain.spawn(fun _ -> send c )) in
  let receivers = Array.init num_domains
                  (fun _ -> Domain.spawn(fun _ -> recv c)) in

  Array.iter Domain.join senders;
  Array.iter Domain.join receivers

(Domain.self () :> int) returns the id of current domain.

Task Distribution using Channels

Now that we have some idea about how channels work, let us consider a more realistic example. We will see how to write a generic task distributor that executes tasks on multiple domains.

module C = Domainslib.Chan
let num_domains = try int_of_string Sys.argv.(1) with _ -> 4
let n = try int_of_string Sys.argv.(2) with _ -> 100

type 'a message = Task of 'a | Quit

let c = C.make_unbounded ()

let create_work tasks =
  Array.iter (fun t -> C.send c (Task t)) tasks;
  for _ = 1 to num_domains do
    C.send c Quit
  done

let rec worker f () =
  match C.recv c with
  | Task a ->
      f a;
      worker f ()
  | Quit -> ()

let _ =
  let tasks = Array.init n (fun i -> i) in
  create_work tasks ;
  let factorial n =
    let rec aux n acc =
        if (n > 0) then aux (n-1) (acc*n)
        else acc in
    aux n 1
  in
  let results = Array.make n 0 in
  let update r i = r.(i) <- factorial i in
  let domains = Array.init (num_domains - 1)
              (fun _ -> Domain.spawn(worker (update results))) in
  worker (update results) ();
  Array.iter Domain.join domains;
  Array.iter (Printf.printf "%d ") results

We have created an unbounded channel c which will act as a store for all the tasks. We'll pay attention to two functions here: create_work and worker.

create_work takes an array of tasks and pushes all elements of tasks to the channel c. The worker function receives tasks from the channel and executes a function f with the received task as parameter. It keeps recursing until it encounters a Quit message, which is why we send Quit messages to the channel, indicating that the worker can terminate.

We can use this template to run any task on multiple cores, by running the worker function on all the domains. This example runs a naive factorial function. The granularity of a task could be tweaked as well, by changing it in the worker function, for instance, worker can run for a range of tasks instead of single one.

Profiling your code

While writing parallel programs in Multicore OCaml, it is quite common to encounter overheads which might deteriorate our code's performance. This section describes ways to discover those overheads and fix them. Linux perf and eventlog in the Multicore runtime are particularly useful tools for performance debugging. In this section, we will be using them for performance debugging. Let's do that with the help of an example.

Perf

Linux perf is a tool that has proved to be very useful to profile Multicore OCaml code.

Profiling serial code

Profiling serial code can help us identify parts of code which can potentially benefit from parallelising. Let's do it for the sequential version of matrix multiplication.

perf record --call-graph dwarf ./matrix_multiplication.exe 1024

We get a profile that tells us how much time is spent in the matrix_multiply function which we wanted to parallelise. What we also need to keep in mind, is that if a lot more time is spent outside the function we'd like to parallelise, the maximum speedup we could achieve would be lower.

Profiling serial code can help us discover the hotspots where we might want to introduce parallelism.

Samples: 51K of event 'cycles:u', Event count (approx.): 28590830181
  Children      Self  Command     Shared Object     Symbol
+   99.84%     0.00%  matmul.exe  matmul.exe        [.] caml_start_program
+   99.84%     0.00%  matmul.exe  matmul.exe        [.] caml_program
+   99.84%     0.00%  matmul.exe  matmul.exe        [.] camlDune__exe__Matmul__entry
+   99.32%    99.31%  matmul.exe  matmul.exe        [.] camlDune__exe__Matmul__matrix_multiply_211
+    0.57%     0.04%  matmul.exe  matmul.exe        [.] camlStdlib__array__init_104
     0.47%     0.37%  matmul.exe  matmul.exe        [.] camlStdlib__random__intaux_278

Overheads in parallel code

Perf can be helpful in identifying overheads in your parallel code. We'll see one such example here where we improve the performance by removing overheads.

Parallel initialisation of a float array with random numbers

Array initialisation using standard library's Array.init is sequential. Parallel workloads in a program would scale according to the number of cores used, whearas the initialisation takes the same amount of time in all cases. This might become a bottleneck for parallel workloads.

For float arrays, we have Array.create_float which creates a fresh float array. We can use this to allocate an array and do the initialisation in parallel. Let us do the initialisation of a float array with random numbers in parallel.

Naive implementation

This is a naive implementation, which will initialise all elements of the array with a Random number.

open Domainslib

let num_domains = try int_of_string Sys.argv.(1) with _ -> 4
let n = try int_of_string Sys.argv.(2) with _ -> 100000
let a = Array.create_float n

let _ =
  let pool = Task.setup_pool ~num_domains:(num_domains - 1) in
  Task.parallel_for pool ~start:0
  ~finish:(n - 1) ~body:(fun i -> Array.set a i (Random.float 1000.));
  Task.teardown_pool pool

Let us measure how it scales.

When we had expected to see speedup executing in multiple cores, what we see here instead is the code slows down as the number of cores increase. There is something wrong with the code which doesn't meet the eye.

We shall profile the performance with perf linux profiler.

$ perf record ./_build/default/float_init_par.exe 4 100_000_000
$ perf report

Perf report would look something like this:

We can see the overhead at Random bits is a whooping 87.99%. Typically there is no one cause that we can attribute to such overheads, since they are very specific to the program. It might need a little careful inspection to find out what is causing them. In this case, the Random module is sharing same state amongst all the domains, which is causing contention when multiple domains are trying to access it at a time.

To overcome that, we shall use a different state for every domain so that there is no contention due to shared state.

module T = Domainslib.Task
let n = try int_of_string Sys.argv.(2) with _ -> 1000
let num_domains = try int_of_string Sys.argv.(1) with _ -> 4

let arr = Array.create_float n

let _ =
  let domains = T.setup_pool ~num_domains:(num_domains - 1) in
  let states = Array.init num_domains (fun _ -> Random.State.make_self_init()) in
  T.parallel_for domains ~start:0 ~finish:(n-1)
  ~body:(fun i ->
    let d = (Domain.self() :> int) mod num_domains in
    Array.unsafe_set arr i (Random.State.float states.(d) 100. ))

We have created num_domains different Random States, each to be used by a different Domain. This might come across as a hack, but if those hacks help us to achieve better performance, there is no harm in using them, as long as the correctness is intact.

We shall run this on multiple cores.

Examining the times, though it is not as bad as the previous case, but it is not close to what we would expect. Let us see the perf report:

The overheads at Random bits is less than the previous case, but it is still quite high at 59.73%. We have used a separate Random State for every domain, so the overheads are not caused by any shared state. But if we look closely, the Random states are all allocated by the same domain in an array with small number of elements, possibly located close to each other in physical memory. When multiple domains try to access them, they might possibly share cache lines, what's termed as false sharing. We can confirm our suspicion with the help of perf c2c on Intel machines.

$ perf c2c record _build/default/float_init_par2.exe 4 100_000_000
$ perf c2c report

Shared Data Cache Line Table     (2 entries, sorted on Total HITMs)
       ----------- Cacheline ----------    Total      Tot  ----- LLC Load Hitm -----  ---- Store Reference ----  --- Loa
Index             Address  Node  PA cnt  records     Hitm    Total      Lcl      Rmt    Total    L1Hit   L1Miss       Lc
    0      0x7f2bf49d7dc0     0   11473    13008   94.23%     1306     1306        0     1560      595      965        ◆
    1      0x7f2bf49a7b80     0     271      368    5.48%       76       76        0      123       76       47        

As evident from the report, there's quite a lot of false sharing happening in the code. To eliminate false sharing, we can allocate the Random state in the domain that is going to use it. This way, the states will be allocated with memory locations far from each other.

module T = Domainslib.Task
let n = try int_of_string Sys.argv.(2) with _ -> 1000
let num_domains = try int_of_string Sys.argv.(1) with _ -> 4

let arr = Array.create_float n

let init_part s e arr =
    let my_state = Random.State.make_self_init () in
    for i = s to e do
      Array.unsafe_set arr i (Random.State.float my_state 100.)
    done

let _ =
  let domains = T.setup_pool ~num_domains:(num_domains - 1) in
  T.parallel_for domains ~chunk_size:1 ~start:0 ~finish:(num_domains - 1)
  ~body:(fun i -> init_part (i * n / num_domains) ((i+1) * n / num_domains - 1) arr);
  T.teardown_pool domains

Now the results are

So, in this process, we have essentially identified bottlenecks for scaling and eliminated them to achieve better speedups. For more details on profiling with perf, please refer these notes.

Eventlog

The Multicore runtime has eventlog enabled by default. Eventlog records GC activity throughout the running of the program. We can generate eventlogs with OCAMLRUNPARAM e:

OCAMLRUNPARAM="e" <executable> <arguments>

This would generate an eventlog in json format. It will be stored in the current working directory. The file can be viewed on chrome://tracing (or brave://tracing). An eventlog file in chrome://tracing looks something like this:

We can zoom in further to find any GC events causing huge latencies.

There is also a script available in ocaml-multicore/tools which displays some statistics from the generated eventlog. It can be invoked as

python3 eventlog_to_latencies.py <eventlog>

This will display some stats like

Mean latency = 33048.10271546635 ns
Max latency = 3590555 ns

## Latency distribution

Percentile, Latency(ns)
10,4887
20,5384
30,5762
40,6298
50,6762
60,7366
70,8401
80,11195
90,15084
95,18888
99,418186
99.9,3587762

## Top slowest events

Latency(ns), Start Timestamp(ns), End TimeStamp(ns), Event, Overhead, Domain ID
3590555, 25535696, 29126251, handle_interrupt, 0, 2
3587762, 25531594, 29119356, dispatch, 0, 3
3360462, 25765906, 29126368, handle_interrupt, 0, 1
3269213, 25925476, 29194689, handle_interrupt, 0, 0
1544112, 11252420, 12796532, dispatch, 0, 3
1539572, 11255922, 12795494, handle_interrupt, 0, 0
1538818, 11256465, 12795283, handle_interrupt, 0, 1
419944, 1064703, 1484647, domain/spawn, 0, 0
418186, 1895401, 2313587, domain/spawn, 0, 0
406474, 1487011, 1893485, domain/spawn, 0, 0
406064, 657246, 1063310, dispatch, 0, 0
199844, 399456161, 399656005, dispatch, 0, 2
197720, 399457883, 399655603, handle_interrupt, 0, 0
197535, 399457879, 399655414, handle_interrupt, 0, 1
185638, 12609615, 12795253, handle_interrupt, 0, 2
76860, 513222912, 513299772, stw/leader, 0, 1
62058, 510995617, 511057675, dispatch, 0, 2
61744, 510994391, 511056135, dispatch, 0, 0
57059, 510998841, 511055900, handle_interrupt, 0, 3
46582, 302806217, 302852799, dispatch, 0, 0
45256, 302807018, 302852274, handle_interrupt, 0, 1
44609, 302807682, 302852291, handle_interrupt, 0, 2
38573, 66232951, 66271524, dispatch, 0, 3
36242, 66235278, 66271520, handle_interrupt, 0, 2
35272, 513264748, 513300020, stw/handler, 0, 3
32050, 19148673, 19180723, dispatch, 0, 3
29817, 305379160, 305408977, dispatch, 0, 1
28746, 513271117, 513299863, stw/handler, 0, 0
27982, 66243537, 66271519, handle_interrupt, 0, 0
27851, 66243628, 66271479, handle_interrupt, 0, 1
27653, 19151760, 19179413, handle_interrupt, 0, 0
27639, 19151690, 19179329, handle_interrupt, 0, 1

We can locate the event which causes maximun latency with the help of the script. Fixing the event may improve the throughput of the program.

Using eventlog to diagnose imbalance in task distribution

Eventlog can be useful to find imbalance in task distribution, if any in a parallel program. Imbalance in task distribution essentially means that, not all domains are provided with equal amount of computation to do. In effect, some domains take longer than others to finish their computations, while the idle domains keep waiting. A possible occurence of this is when a sub- optimal chunk_size is picked in a parallel_for.

Time periods when a domain is idle is recorded as domain/idle_wait in the eventlog. Here is an example eventlog generated by a program with unbalanced task distribution.

If we zoom in further, we see many domain/idle_wait events.

So far we have only found that there is an imbalance in task distribution in the code, we'll need to change our code accordingly to make the task distribution more balanced, which could increase the speedup.

Performace debugging can be quite tricky at times. If you could use some help in debugging your Multicore OCaml code, feel free to create an issue in the Multicore OCaml issue tracker along with a minimal code example.