Benchmark history for the async primitives in CryptoHives.Foundation.Threading,
measured against Nito.AsyncEx, ProtoPromise, Microsoft.VisualStudio.Threading, NeoSmart.AsyncLock,
DotNext.Threading, the keyed-lock libraries (AsyncKeyedLock, KeyedSemaphores, Dao.IndividualLock,
AsyncUtilities) and the BCL primitives they replace.
Pick a platform, family and method; each implementation is its own series. Everything runs in your browser β no server, no external database.
Windows runs from 29 September 2026 on cap the maximum processor state at 99%, which keeps the CPU out of its boost clocks. Results are steadier from run to run, but about 10% slower across the board β a step down in a Windows trend line at that date is the measurement setup, not a regression.
Click a color, line style or point shape to hide or show that series.
Table (one run) reproduces one run's report as BenchmarkDotNet wrote it. Our own rows are shaded, the declared baseline more deeply; Ratio is relative to that baseline, which for some families falls outside the current selection. Pick any run, and optionally a second to compare β runs of the same commit on two machines are marked (same commit).
Trend over time plots one point per recorded run for every parameter combination; the line-style legend hides combinations. Scaling by contention plots the parameterized axis (queued waiters, or distinct keys/threads for the keyed locks) for one run: color is the implementation, line style the cancellation state, mean time on the left axis and allocation on the right.
Hover a point for the commit, the value, the .NET runtime and the reference library's version.
A live but unfired CancellationToken has measurable overhead over
CancellationToken.None, which is why Cancellation is a selector of its own.
Data: benchmark-history.sqlite, built from the benchmarks branch by
scripts/threading-benchmark-trends/import_run_archive.py.