Context Swap Microbenchmark - #381

Open

insertinterestingnamehere wants to merge 21 commits into

sandialabs:mainfrom

insertinterestingnamehere:bench

Commits on Aug 19, 2026

Add a microbenchmark for testing context swapping overhead.
insertinterestingnamehere
committed
3a250cf
Fix using the test argparsing header from C++.
insertinterestingnamehere
committed
03ce306
Add minimal syscall benchmark.
insertinterestingnamehere
committed
e83a6db
Add os thread fork/join benchmark.
insertinterestingnamehere
committed
acbc337
Add os_thread_futex_handoff benchmark.
insertinterestingnamehere
committed
13c7949
Add benchmark measuring function call overheads for various numbers of arguments.
insertinterestingnamehere
committed
c6aed48
Require C++17 for building tests.
insertinterestingnamehere
committed
0591808
Add a benchmark that attempts to force branch mispredicted function calls.
insertinterestingnamehere
committed
107e1f5
Rewrite function call overhead benchmark to not require a bunch of horrible manual repetition.
insertinterestingnamehere
committed
6578d25
Add mini nemesis many-to-one benchmark.
insertinterestingnamehere
committed
870f8c6
Restructure ifdefs to allow using c++23 or later in the benchmarks.
insertinterestingnamehere
committed
870be7d
Make the cast to volatile explicit in the function call microbenchmark to avoid clang warnings.
insertinterestingnamehere
committed
5a8fef3
Use C++26 for the benchmark suite. The benchmarks aren't required for the main library, so we can have nice things here.
insertinterestingnamehere
committed
c1e3cc1
Make qt_atomics usable from within C++ and use that directly to get the platform-generic pause macro inside the benchmarks.
insertinterestingnamehere
committed
e82d1b0
Add benchmarks for a few different atomic workloads.
insertinterestingnamehere
committed
601015c
Add contended atomic exchange benchmark.
insertinterestingnamehere
committed
8ab36e2
Fix race condition in mini nemesis queue benchmark.
insertinterestingnamehere
committed
a5d8960
Switch nemesis_many_to_one benchmark to specify the total number of work items instead of the number per thread. This makes scalability testing easier.
insertinterestingnamehere
committed
7f28e19
Revive pure qthreads thread-ring benchmark.
insertinterestingnamehere
committed
dfd2fc4
Fix typo in CI job naming.
insertinterestingnamehere
committed
89dadf0
Make gcc16 happy with the atomic fetch_sub benchmark by using a fence instead of relying on volatile.
insertinterestingnamehere
committed
f8b2e9b

Navigation Menu

Uh oh!

Context Swap Microbenchmark - #381

Context Swap Microbenchmark#381
insertinterestingnamehere wants to merge 21 commits into
sandialabs:mainfrom
insertinterestingnamehere:bench

Commits on Aug 19, 2026

Add a microbenchmark for testing context swapping overhead.

Fix using the test argparsing header from C++.

Add minimal syscall benchmark.

Add os thread fork/join benchmark.

Add os_thread_futex_handoff benchmark.

Add benchmark measuring function call overheads for various numbers of arguments.

Require C++17 for building tests.

Add a benchmark that attempts to force branch mispredicted function calls.

Rewrite function call overhead benchmark to not require a bunch of horrible manual repetition.

Add mini nemesis many-to-one benchmark.

Restructure ifdefs to allow using c++23 or later in the benchmarks.

Make the cast to volatile explicit in the function call microbenchmark to avoid clang warnings.

Use C++26 for the benchmark suite. The benchmarks aren't required for the main library, so we can have nice things here.

Make qt_atomics usable from within C++ and use that directly to get the platform-generic pause macro inside the benchmarks.

Add benchmarks for a few different atomic workloads.

Add contended atomic exchange benchmark.

Fix race condition in mini nemesis queue benchmark.

Switch nemesis_many_to_one benchmark to specify the total number of work items instead of the number per thread. This makes scalability testing easier.

Revive pure qthreads thread-ring benchmark.

Fix typo in CI job naming.

Make gcc16 happy with the atomic fetch_sub benchmark by using a fence instead of relying on volatile.