Landmarks: A Simple Profiling Library

Landmarks is a simple profiling library for OCaml. It provides primitives to delimit portions of code and measure the performance of instrumented code at runtime. The available measures are obtained by aggregating CPU cycles (using the cpu's time stamp counter), applicative time (using Sys.time) and allocated bytes (with Gc.allocated_bytes). The instrumentation of the code may either done by hand, automatically or semi-automatically using a PPX extension.

During the execution of your program, the traversal of instrumented code by the control flow is recorded as a "callgraph" that carries the collected measures. The results may be browsed either directly on the console, or exported to json.

This tool is intended to be used as a way to find where the time is spent in your programs (and not benchmark independent pieces of code like Core_bench) while providing results that only correspond to the instrumented portion of your OCaml code (contrary to tools that directly work with the binary executable like gprof or perf).

For more information, you may browse the API.

Installation

The library is split into two packages: landmarks for the runtime library and landmarks-ppx for the preprocessor implementing automatic instrumentation.

• Using opam:

opam install landmarks

or

opam install landmarks-ppx

will install either the runtime library or both the runtime library and the preprocessor.

• Manually:

git clone https://github.com/LexiFi/landmarks.git
cd landmarks
dune build @install

Usage with dune

Direct Method

Simply use the library landmarks and the preprocessor landmarks-ppx to benchmark your executables and libraries. For instance, the following dune file builds the executable test using the landmarks library and its PPX. The optional --auto flag turns on the automatic instrumentation (see below).

(executable
 (name test)
 (libraries landmarks)
 (preprocess (pps landmarks-ppx --auto))
)

Using dune's instrumentation

It is possible to use dune to automatically trigger the instrumentation of a project. Have a look at LexiFi/landmarks-starter for a basic example and see the dune manual for more information.

Usage with ocamlfind

• Compiling and linking:

  ocamlfind ocamlopt -c -package landmarks prog.ml
  ocamlfind ocamlopt -o prog -package landmarks -linkpkg prog.cmx

You can replace "ocamlopt" by "ocamlc" to compile the program in bytecode.

• With the PPX extension:

  ocamlfind ocamlopt -c -package landmarks -package landmarks-ppx -ppxopt landmarks-ppx,--auto prog.ml
  ocamlfind ocamlopt -o prog -package landmarks -linkpkg prog.cmx

Note that "-ppxopt landmarks-ppx,--auto" is optional and turns on the automatic instrumentation.

Benchmarking manually

There are three main primitives:

  val register: string -> landmark
  val enter: landmark -> unit
  val exit: landmark -> unit

The register function declares new landmarks and should be used at the toplevel. The functions enter and exit are used to delimit the portion of code attached to a landmark. At the end of the profiling, we retrieve for each landmark the aggregated time information spent executing the corresponding piece of code. During the execution, a trace of each visited landmark is also recorded in order to build a "callgraph".

For example:

open Landmark

let loop = register "loop"
let sleep = register "sleep"
let main = register "main"

let zzz () =
  enter sleep;
    Unix.sleep 1;
  exit sleep

let () =
  begin
    start_profiling ();
    enter main;
      enter loop;
        for _ = 1 to 9 do
          zzz ()
        done;
      exit loop;
      zzz ();
    exit main;
  end

(This file can be compiled with ocamlfind ocamlc -o prog -package landmarks -package unix -linkpkg prog.ml)

The induced callgraph is:

- 100.00% : main
|   - 90.00% : loop
|   |   - 100.00% : sleep
|   - 10.00% : sleep

which can be paraphrased as:

• 100% of time is spent inside the main landmark,
• 90% of time spent inside the main landmark is spent in the loop landmark,
• 10% of time spent inside the main landmark is spent in the sleep landmark,
• 100% of the time spent in loop is spent in the sleep landmark.

The clock() function

The library provides a binding to the high-performance cycles counter for x86 32 and 64 bits architectures (note that you may use the landmarks-noc.cm(x)a archive to provide your own implementation). It is used to measure the time spent inside instrumented code.

The PPX extension point

To avoid writing boilerplate code, you may use the ppx extension distributed with this package. It allows the programmer to instrument expressions using annotation and to automatically instrument top-level functions.

Annotations

The value expr [@landmark "name"] is expanded into

  Landmark.enter __generated_landmark_1;
  let r =
    try expr with e -> Landmark.exit __generated_landmark_1; raise e
  in
  Landmark.exit __generated_landmark_1;
  r

and the declaration

  let __generated_landmark_1 = Landmark.register "name"

is appended at the top-level.

It should be pointed out that this transformation does not preserve tail-recursive calls (and also prevents some polymorphism generalization). To get around these problems, it is recommended to use the other provided extension around let ... in and let rec ... in:

let[@landmark] f = body

which is expanded in :

let __generated_landmark_2 = Landmark.register "f"
let f = body
let f x1 ... xn =
  Landmark.enter __generated_landmark_2;
  let r =
    try f x1 ... xn with e -> Landmark.exit __generated_landmark_2; raise e
  in
  Landmark.exit __generated_landmark_2;
  r

when the arity n of f is obtained by counting the shallow occurrences of fun ... -> and function ... -> in body. Please note that when using this annotation with let-rec bindings, only entry-point calls will be recorded. For instance, in the following piece of code

  let () =
    let[@landmark] rec even n = (n = 0) || odd (n - 1)
    and[@landmark] odd n = (n = 1) || n > 0 && even (n - 1)
    in Printf.printf "'six is even' is %b\n" (even 6)

the landmark associated with "even" will be traversed exactly once (and not three times !) whereas the control flow will not pass through the landmark associated with "odd".

Automatic instrumentation

The structure annotations [@@@landmark "auto"] and [@@@landmark "auto-off"] activate or deactivate the automatic instrumentation of top-level functions in a module. In automatic mode, all functions declarations are implicitly annotated. Automatic instrumentation can be enabled/disabled for all files via option auto in OCAML_LANDMARKS, as detailed below.

The OCAML_LANDMARKS environment variable

The environment variable OCAML_LANDMARKS is read at two different stages: when the ppx rewriter is executed, and when the landmarks module is loaded by an instrumented program. This variable is parsed as a comma-separated list of items of the form option=argument or option, where option is:

• During the ppx rewriter stage (at compilation time):

  • auto (no arguments): turns on the automatic instrumentation by default (behaves as if each module starts with annotation [@@@landmark "auto"]).

  • threads (no arguments): tells the ppx extension to use the Landmark_threads module instead of the module Landmark.

• When loading an instrumented program (at runtime):

  • format with possible arguments: textual (default) or json. It controls the output format of the profiling which is either a console friendly representation or json encoding of the callgraph.

  • threshold with a number between 0.0 and 100.0 as argument (default: 1.0). If the threshold is not zero the textual output will hide nodes in the callgraph below this threshold (in percent of time of their parent). This option is meaningless for other formats.

  • output with possible argument: stderr (default), stdout, temporary, <file> (where <file> is the path of a file). It tells where to output the results of the profiling. With temporary it will print it in a temporary file (the name of this file will be printed on the standard error). You may also use temporary:<directory> to specify the directory where the files are generated.

  • debug with no argument. Activates a verbose mode that outputs traces on stderr each time the landmarks primitives are called.

  • time with no argument. Also collect Sys.time timestamps during profiling.

  • off with no argument. Disable profiling.

  • on with no argument. Enable profiling (default; may be omitted).

  • allocation with no argument. Also collect Gc.allocated_byte data.

Browsing the JSON export using the Web Viewer

You can either compile the web viewer on your computer or browse it online. You need to load the JSON files using the filepicker and then you can click around to browse the callgraph.

Instrumenting with threads

The Landmark module is not thread-safe. If you have multiple threads, you have to make sure that at most one thread is executing instrumented code. For that you may use the Landmark_threads module (included in the landmarks-threads.cm(x)a archive) that prevents non thread-safe functions to execute in all threads but the one which started the profiling.

Known Issue

The annotation on expressions may temper with polymorphism (this is not the case for the let-binding annotation). For instance, the following piece of code will fail to compile:

  let test = (fun x -> x)[@landmark "test"]
  in test "string", test 1

About

This 'Landmarks' package is licensed by LexiFi under the terms of the MIT license.

Contact: [email protected]