How to run

From Visual Studio

Ensure you're in Release configuration, then run without debugging (Ctrl + F5).

From the command line

dotnet run -c Release

Filtering out benchmarks

By default, Benchmark.NET runs every benchmark it finds in the executable. To run specific benchmarks, you can specify the --filter argument at the command line.

For example, to run only call benchmarks you can run

dotnet run -c Release --filter *CallBench*

Call Benchmark

This benchmark compares synchronous, asynchronous, and mixed method calls.

AsyncCall is an async method that awaits a call to another async method. MixedCall is an async method that awaits a call to a synchronous method. SyncCall is a synchronous method that directly calls another synchronous method.

Benchmark results

Method	Mean	Error	StdDev
AsyncCall	35.992 ns	0.5083 ns	0.4506 ns
MixedCall	21.674 ns	0.1605 ns	0.1423 ns
FullSyncCall	4.416 ns	0.1147 ns	0.1073 ns

Async Lazy Benchmark

Compares awaiting a task directly and awaiting a task wrapped by AsyncLazy<T>.

Benchmark results

Method	Mean	Error	StdDev
AsyncLazySingle	1,537.24 ns	30.529 ns	63.726 ns
TaskSingle	34.76 ns	0.449 ns	0.420 ns
AsyncLazyMulti	1,895.75 ns	14.187 ns	13.270 ns
YieldingAsyncLazySingle	2,354.20 ns	10.358 ns	9.689 ns
YieldingTaskSingle	1,341.59 ns	4.837 ns	4.525 ns
YieldingAsyncLazyMulti	2,844.05 ns	23.930 ns	22.385 ns

Repository Benchmarks

AllEagerProps benchmarks eagerly load Account properties and read all of them. AllLazyProps benchmarks lazily load Account properties and read all of them.

TwoEagerProps benchmarks eagerly load Account properties and read 2 of them. TwoLazyProps benchmarks lazily load Account properties and read 2 of them.

UniqueAccount benchmarks read from 100 unique accounts. ReusedAccount benchmarks read from 1 account, but 5 separate times.

Async Repo Benchmark

This benchmark simulates a call to repository backed by some form of external storage. The AsyncRepo simulates an external call by awaiting Task.Yield() - this doesn't simulate the delay of a database call, but it should be fairly accurate on the overhead of async/await itself.

Benchmark results

Method	Mean	Error	StdDev
AllEagerPropsUniqueAccount	601.037 us	1.8149 us	1.5155 us
AllLazyPropsUniqueAccount	1,250.706 us	7.2732 us	6.4475 us
TwoEagerPropsUniqueAccount	600.131 us	3.3568 us	3.1399 us
TwoLazyPropsUniqueAccount	517.754 us	2.5689 us	2.4030 us

Method	Mean	Error	StdDev
AllEagerPropsReusedAccount	6.233 us	0.0329 us	0.0291 us
AllLazyPropsReusedAccount	17.243 us	0.1381 us	0.1292 us
TwoEagerPropsReusedAccount	6.224 us	0.0396 us	0.0331 us
TwoLazyPropsReusedAccount	5.477 us	0.0381 us	0.0338 us

Sync Repo Benchmark

These benchmarks simulate a call to a repository that doesn't need to go out to external storage, such as when retrieving something that's cached, or a hardcoded database table.

The benchmarks here should always be faster than the Async Repo benchmarks, as the underlying repository does no awaiting or yielding, instead simply returning a completed task.

Benchmark results:

Method	Mean	Error	StdDev
AllEagerPropsUniqueAccount	7,549.12 ns	92.948 ns	86.944 ns
AllLazyPropsUniqueAccount	683,726.47 ns	4,633.499 ns	4,334.178 ns
TwoEagerPropsUniqueAccount	7,238.91 ns	62.664 ns	58.616 ns
TwoLazyPropsUniqueAccount	276,023.49 ns	3,258.254 ns	2,888.357 ns

Method	Mean	Error	StdDev
AllEagerPropsReusedAccount	87.15 ns	0.644 ns	0.602 ns
AllLazyPropsReusedAccount	7,411.52 ns	55.369 ns	51.792 ns
TwoEagerPropsReusedAccount	86.33 ns	1.657 ns	1.384 ns
TwoLazyPropsReusedAccount	3,229.06 ns	18.189 ns	17.014 ns

Continue Benchmark

Compares ContinueWith and await

Benchmark results:

Method	Mean	Error	StdDev
AwaitYields	4,459.67 ns	81.541 ns	97.069 ns
ContinueWithYields	8,088.39 ns	157.621 ns	258.977 ns
AwaitCompleted	86.09 ns	0.540 ns	0.451 ns
ContinueWithCompleted	7,269.94 ns	58.798 ns	54.999 ns
AwaitDelay	29,869,587.99 ns	594,620.063 ns	1,214,651.218 ns
ContinueWithDelay	27,881,613.52 ns	554,491.159 ns	1,027,784.916 ns