#benchmark #benchmarking #performance

app sightglass

A benchmark suite and tool to compare different implementations of the same primitives

1 unstable release

0.1.0 Apr 8, 2019

#205 in Profiling

Apache-2.0 WITH LLVM-exception OR MIT

21K SLoC

C 20K SLoC // 0.0% comments Rust 1K SLoC // 0.0% comments JavaScript 56 SLoC // 0.2% comments Vue 17 SLoC Stylus 3 SLoC

Sight Glass

A benchmark suite and tool to compare different implementations of the same primitives.


Sight Glass loads multiple shared libraries implementing the same test suite, runs all tests from all suites, and produces reports to evaluate how implementations compare to each other.

Functions from each library are evaluated as follows:


  test1_setup(global_ctx, &test1_ctx);

  test2_setup(global_ctx, &test2_ctx);

  // ...

  testN_setup(global_ctx, &testN_ctx);


Each shared library must export a tests_config symbol:

typedef struct TestsConfig {
    void     (*global_setup)(void **global_ctx_p);
    void     (*global_teardown)(void *global_ctx);
    uint64_t version;
} TestsConfig;

TestsConfig tests_config;

global_setup and global_teardown are optional, and can be set to NULL if not required.

A test must at least export a function named <testname>_body:

void testname_body(void *ctx);

This function contains the actual code to be benchmarked.

By default, ctx will be set to the global_ctx. However, optional setup and teardown functions can also be provided for individual tests:

void testname_setup(void *global_ctx, void **ctx_p);

void testname_teardown(void *ctx);

See example/example.c for an example test suite.

Sightglass extracts all symbols matching the above convention to define and run the test suite.

Running multiple functions for a single test

A single test can evaluate multiple body functions sharing the same context.

These functions have to be named <testname>_body_<bodyname>.

<bodyname> can be anything; a numeric ID or a short description of the purpose of the function.

void testname_body_2(void *ctx);
void testname_body_randomized(void *ctx);

These functions are guaranteed to be evaluated according to their names sorted in lexical order.


The global configuration is loaded from sightglass.toml file. This can be changed using the -c command-line flag.

The configuration lists implementations to be benchmarked:

test_suites = [
  { name = "test1", library_path = "implementation1.so" },
  { name = "test2", library_path = "implementation2.so" }

Individual test suites can also run a command in order to be optionally skipped if that command returns a non-zero exit code:

test_suites = [
  { name = "test1", library_path = "implementation1.so" },
  { name = "test2", library_path = "implementation2.so", guard = ["/opt/sg/guard-scripts/check", "arg1", "arg2"] }

Additional properties that the file can include:

  • single_core = <bool>: set to true in order to run the tests on a single CPU core, in order to get more accurate results. This only works on Linux.

  • output = [ { format = "Text|CSV|JSON" [, file = <file>] [, breakdown = <bool>] } ... ]: how to store or display the results.

By defaut, the Text and CSV output do not include a breakdown of the time spent in individual functions for tests made of multiple functions. This can be changed with the optional breakdown property being set to true.

The JSON output always includes this information.


~69K SLoC