thanks to Nico Werner, who did most of the porting work
12 KiB
Benchmarks
The benchmarks are done with this script using CMake. There are 3 benchmarking scenarios:
Compilers used:
- WINDOWS: Microsoft Visual Studio Community 2017 - Version 15.8.1+28010.2003
- WINDOWS: gcc 8.1.0 (x86_64-posix-seh-rev0, Built by MinGW-W64 project)
- LINUX: gcc 6.3.0 20170406 (Ubuntu 6.3.0-12ubuntu2)
- LINUX: clang 4.0.0-1 (tags/RELEASE_400/rc1) Target: x86_64-pc-linux-gnu
Environment used (Intel i7 3770k, 16g RAM):
- Windows 7 - on an SSD
- Ubuntu 17.04 in a VirtualBox VM - on a HDD
doctest version: 2.2.0 (released on 2018.12.02)
Catch version: 2.3.0 (released on 2018.07.22)
Compile time benchmarks
Cost of including the header
This is a benchmark that is relevant only to single header and header only frameworks - like doctest and Catch.
The script generates 201 source files and in 200 of them makes a function in the form of int f135() { return 135; }
and in main.cpp
it forward declares all the 200 such dummy functions and accumulates their result to return from the main()
function. This is done to ensure that all source files are built and that the linker doesn't remove/optimize anything.
- baseline - how much time the source files need for a single threaded build with
msbuild
/make
- + implement - only in
main.cpp
the header is included with a#define
before it so the test runner gets implemented:
#define DOCTEST_CONFIG_IMPLEMENT_WITH_MAIN
#include "doctest.h"
- + header everywhere - the framework header is also included in all the other source files
- + disabled - remove everything testing-related from the binary
doctest | baseline | + implement | + header everywhere | + disabled |
---|---|---|---|---|
MSVC Debug | 4.89 | 6.21 | 8.33 | 6.39 |
MSVC Release | 4.38 | 6.39 | 8.71 | 6.02 |
MinGW GCC Debug | 8.12 | 10.86 | 14.73 | 10.17 |
MinGW GCC Release | 8.21 | 11.11 | 15.03 | 10.71 |
Linux GCC Debug | 4.20 | 6.23 | 9.81 | 6.24 |
Linux GCC Release | 4.29 | 6.93 | 11.05 | 6.76 |
Linux Clang Debug | 8.70 | 10.02 | 14.43 | 11.13 |
Linux Clang Release | 9.30 | 11.68 | 16.20 | 11.58 |
Catch | baseline | + implement | + header everywhere | + disabled |
---|---|---|---|---|
MSVC Debug | 4.82 | 7.83 | 88.85 | 88.72 |
MSVC Release | 4.38 | 9.97 | 87.17 | 88.35 |
MinGW GCC Debug | 8.00 | 57.28 | 137.28 | 132.73 |
MinGW GCC Release | 8.38 | 22.94 | 97.17 | 97.22 |
Linux GCC Debug | 4.42 | 15.57 | 97.94 | 97.18 |
Linux GCC Release | 4.50 | 19.59 | 99.48 | 100.75 |
Linux Clang Debug | 8.76 | 15.60 | 107.99 | 110.61 |
Linux Clang Release | 9.32 | 25.75 | 118.67 | 117.11 |
Conclusion
doctest
- instantiating the test runner in one source file costs ~1-3 seconds
implement - baseline
- the inclusion of
doctest.h
in one source file costs between 11ms - 23ms(header_everywhere - implement) / 200
- including the library everywhere but everything disabled costs around 2 seconds
disabled - baseline
for 200 files
Catch
- instantiating the test runner in one source file costs ~3-50 seconds
implement - baseline
- the inclusion of
catch.hpp
in one source file costs between 380ms - 470ms(header_everywhere - implement) / 200
- using the config option to disable the library (
CATCH_CONFIG_DISABLE
) has no effect on the header cost
So if doctest.h
costs 11ms and catch.hpp
costs 400ms on MSVC - then the doctest header is >> 36 << times lighter (for MSVC)!
The results are in seconds and are in no way intended to bash Catch - the doctest framework wouldn't exist without it.
The reason the doctest header is so light on compile times is because it forward declares everything and doesn't drag any headers in the source files (except for the source file where the test runner gets implemented). This was a key design decision.
Cost of an assertion macro
The script generates 11 .cpp
files and in 10 of them makes 50 test cases with 100 asserts in them (of the form CHECK(a==b)
where a
and b
are always the same int
variables) - 50k asserts! The testing framework gets implemented in main.cpp
.
- baseline - how much time a single threaded build takes with the header included everywhere - no test cases or asserts!
CHECK(a==b)
- will addCHECK()
asserts which decompose the expression with template machinery
doctest specific:
- +fast 1 - will add
DOCTEST_CONFIG_SUPER_FAST_ASSERTS
to speed up the compilation of the normal assertsCHECK(a==b)
CHECK_EQ(a,b)
- will useCHECK_EQ(a,b)
instead of the expression decomposing ones- +fast 2 - will add
DOCTEST_CONFIG_SUPER_FAST_ASSERTS
to speed up the compilation of the binary assertsCHECK_EQ(a,b)
- +disabled - all test case and assert macros will be disabled with
DOCTEST_CONFIG_DISABLE
Catch specific:
- +fast - will add
CATCH_CONFIG_FAST_COMPILE
which speeds up the compilation of the normal assertsCHECK(a==b)
- +disabled - all test case and assert macros will be disabled with
CATCH_CONFIG_DISABLE
doctest | baseline | CHECK(a==b) |
+fast 1 | CHECK_EQ(a,b) |
+fast 2 | +disabled |
---|---|---|---|---|---|---|
MSVC Debug | 2.69 | 27.37 | 10.37 | 17.17 | 4.82 | 1.91 |
MSVC Release | 3.15 | 58.73 | 20.73 | 26.07 | 6.43 | 1.83 |
MinGW GCC Debug | 3.78 | 97.29 | 43.05 | 59.86 | 11.88 | 1.67 |
MinGW GCC Release | 4.09 | 286.70 | 95.42 | 156.73 | 18.16 | 2.03 |
Linux GCC Debug | 2.39 | 91.36 | 41.92 | 52.26 | 10.16 | 1.32 |
Linux GCC Release | 3.29 | 257.40 | 97.46 | 128.84 | 19.38 | 1.79 |
Linux Clang Debug | 2.40 | 85.52 | 43.53 | 51.24 | 8.32 | 1.62 |
Linux Clang Release | 3.40 | 160.65 | 79.34 | 81.52 | 11.90 | 1.82 |
And here is Catch which only has normal CHECK(a==b)
asserts:
Catch | baseline | CHECK(a==b) |
+fast | +disabled |
---|---|---|---|---|
MSVC Debug | 8.20 | 31.22 | 25.54 | 8.22 |
MSVC Release | 10.13 | 448.68 | 168.67 | 10.20 |
MinGW GCC Debug | 53.54 | 152.38 | 131.85 | 49.07 |
MinGW GCC Release | 19.26 | 590.16 | 466.69 | 18.99 |
Linux GCC Debug | 15.05 | 117.30 | 95.33 | 14.79 |
Linux GCC Release | 18.77 | 608.94 | 482.73 | 18.96 |
Linux Clang Debug | 12.27 | 94.39 | 77.33 | 12.11 |
Linux Clang Release | 20.75 | 545.84 | 506.02 | 20.15 |
Conclusion
doctest:
- is between 0 and 8 times faster than Catch when using normal expression decomposing
CHECK(a==b)
asserts - asserts of the form
CHECK_EQ(a,b)
with no expression decomposition - around 31-63% faster thanCHECK(a==b)
- the
DOCTEST_CONFIG_SUPER_FAST_ASSERTS
identifier makes the normal asserts faster by 57-68% - the
DOCTEST_CONFIG_SUPER_FAST_ASSERTS
identifier makes the binary asserts even faster by another 84-91% - using the
DOCTEST_CONFIG_DISABLE
identifier the asserts just disappear as if they were never written - even lower than the baseline (because most of the implementation is also gone)
- using
CATCH_CONFIG_FAST_COMPILE
results in 10-30% faster build times for asserts (and in one case 73%). - using the
CATCH_CONFIG_DISABLE
identifier provides the same great benefits for assert macros as the doctest version (DOCTEST_CONFIG_DISABLE
) - but not for the header cost
Runtime benchmarks
The runtime benchmarks consist of a single test case with a loop of 10 million iterations performing the task - a single normal assert (using expression decomposition) or the assert + the logging of the loop iterator i
:
for(int i = 0; i < 10000000; ++i)
CHECK(i == i);
or
for(int i = 0; i < 10000000; ++i) {
INFO(i);
CHECK(i == i);
}
Note that the assert always passes - the goal should be to optimize for the common case - lots of passing test cases and a few that maybe fail.
doctest | assert | + info | Catch | assert | + info | |
---|---|---|---|---|---|---|
MSVC Debug | 4.00 | 11.41 | MSVC Debug | 5.60 | 213.91 | |
MSVC Release | 0.40 | 1.47 | MSVC Release | 0.76 | 7.60 | |
MinGW GCC Debug | 1.05 | 2.93 | MinGW GCC Debug | 1.17 | 9.54 | |
MinGW GCC Release | 0.34 | 1.27 | MinGW GCC Release | 0.36 | 4.28 | |
Linux GCC Debug | 1.24 | 2.34 | Linux GCC Debug | 1.44 | 9.69 | |
Linux GCC Release | 0.29 | 0.52 | Linux GCC Release | 0.29 | 3.60 | |
Linux Clang Debug | 1.15 | 2.38 | Linux Clang Debug | 1.21 | 9.91 | |
Linux Clang Release | 0.28 | 0.50 | Linux Clang Release | 0.32 | 3.27 |
Conclusion
doctest is around ~20% faster than catch for asserts but a few times faster when also logging variables and context (and in the case of one particular compiler over 18 times faster).
The bar charts were generated using this google spreadsheet by pasting the data from the tables.
If you want a benchmark that is not synthetic - check out this blog post of Baptiste Wicht who tested the compile times of the asserts in the 1.1 release with his Expression Templates Library!
While reading the post - keep in mind that if a part of a process takes 50% of the time and is made 10000 times faster - the overall process would still be only roughly 50% faster.