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.
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
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))
)
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.
- 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.
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 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.
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.
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".
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 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 theLandmark_threads
module instead of the moduleLandmark
.
-
-
When loading an instrumented program (at runtime):
-
format
with possible arguments:textual
(default) orjson
. 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. Withtemporary
it will print it in a temporary file (the name of this file will be printed on the standard error). You may also usetemporary:<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 collectSys.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 collectGc.allocated_byte
data.
-
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.
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.
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
This 'Landmarks' package is licensed by LexiFi under the terms of the MIT license.
Contact: marc.lasson@lexifi.com