Benchmarks

All numbers come from BenchmarkDotNet with MemoryDiagnoser, on real hardware. No synthetic inflation.

What these measure

Framework dispatch overhead with no-op handlers — the handler returns immediately. That isolates the mediator’s own cost. Real handlers doing I/O or CPU work narrow the relative gap: the absolute nanosecond and allocation differences remain, but they become a smaller fraction of total request time. Numbers are from a single fresh run; expect a few percent of run-to-run variance.

Environment

Property	Value
Runtime	.NET 10.0.1, Arm64 RyuJIT AdvSIMD
OS / CPU	macOS, Apple M5
BenchmarkDotNet	v0.14.0
MediatR	v12.4.1
Mediant	v1.0.0

Publish — notification fanout

Where Mediant wins decisively. MediatR allocates a NotificationHandlerExecutor wrapper plus a closure delegate per handler, per call. Mediant invokes handlers directly — zero wrapper allocation.

Handlers	Mediant	MediatR v12	Speed	Memory
1	24 ns / 88 B	44 ns / 288 B	46% faster	3.3× less
10	69 ns / 376 B	184 ns / 1,656 B	63% faster	4.4× less
50	276 ns / 1,656 B	792 ns / 7,736 B	65% faster	4.7× less
100	527 ns / 3,256 B	1,521 ns / 15,336 B	65% faster	4.7× less

Send — command/query pipeline

The hot path: every request goes through Send(). Behaviors are scaled exponentially to show how each library handles pipeline depth.

Behaviors	Mediant	MediatR v12	Speed	Memory
0	24 ns / 64 B	24 ns / 128 B	~equal	2× less
1	59 ns / 288 B	57 ns / 368 B	~equal	22% less
2	75 ns / 424 B	70 ns / 512 B	~equal	17% less
4	102 ns / 696 B	100 ns / 800 B	~equal	13% less
8	150 ns / 1,240 B	159 ns / 1,376 B	6% faster	10% less
16	262 ns / 2,328 B	277 ns / 2,528 B	5% faster	8% less
32	478 ns / 4,504 B	507 ns / 4,832 B	6% faster	7% less

Speed is within noise through 4 behaviors; from 8 up Mediant pulls ahead, and the gap widens with depth. Memory is lower in every scenario.

Query — with a return value

Scenario	Mediant	MediatR v12	Speed	Memory
Query returning Result<int>	28 ns / 104 B	28 ns / 200 B	~equal	1.9× less

Note that Mediant returns the richer Result<int> here, while MediatR returns a raw int.

Scorecard

Category	Benchmarks	Mediant wins	Tie	MediatR wins
Send (8+ behaviors)	3	3	0	0
Send (0–4 behaviors)	4	0	4	0
Query	1	0	1	0
Publish	4	4	0	0
Total	12	7	5	0

Mediant allocates less memory in all 12 benchmarks, and never loses on speed.

Load tests

18 load tests, part of a 300-test suite spanning unit, integration, load and E2E.

Concurrency and throughput

Test	Scale	Result
Concurrent Send	10,000 simultaneous	No deadlocks, all succeed
Concurrent Send + behaviors	50,000 simultaneous	No deadlocks, all succeed
Concurrent Query	5,000 simultaneous	All succeed, correct results
Scoped DI (per request)	10,000 scopes	All independent
Mixed operations	10,000	All complete cleanly

Memory and stability

Test	Scale	Result
Sequential memory leak	100,000 requests	< 10 MB growth
Sequential + behaviors	500,000 requests	< 20 MB growth
High-cardinality cache keys	10,000 unique keys	< 20 MB growth
Thread-pool exhaustion	20,000 operations	Pool not depleted

Resilience

Test	Scale	Result
Exceptions under load	10,000 (33% failing)	All complete, no leaks
Cancellation mid-flight	5,000 + cancel	No hanging tasks
Re-entrant Send	1,000 × depth 3	No deadlocks
Parallel fanout	5,000 × 10 handlers	50,000 executions, all succeed
Streaming consumers	100 × 1,000 items	All correct

Sustained latency (10 seconds): p50 < 1 ms, p95 < 5 ms, p99 < 10 ms, throughput > 10,000 req/sec.

Reproduce them

git clone https://github.com/omercelikdev/mediant.git
cd mediant

# BenchmarkDotNet comparison vs MediatR
dotnet run -c Release --project tests/Mediant.Benchmarks

# Load tests
dotnet test tests/Mediant.LoadTests