TurboMqtt MQTT 5.0 Performance Benchmarks

February 22, 2026 · View on GitHub

Full production benchmarks for the MQTT 5.0 pipeline, recorded as part of the pre-1.0 release quality gate (GitHub issue #371).

Machine: Intel Core i9-9900K CPU 3.60GHz (Coffee Lake), 1 CPU, 8 logical and 8 physical cores OS: Linux Ubuntu 24.04.4 LTS (Noble Numbat) Runtime: .NET 10.0.3, X64 RyuJIT x86-64-v3 BenchmarkDotNet: v0.15.8, RunStrategy=Monitoring, LaunchCount=10, WarmupCount=10 Benchmark source: Mqtt5End2EndTcpBenchmarks.cs, Mqtt5TlsTcpBenchmarks.cs

Each benchmark iteration publishes and receives 1,000 messages end-to-end through an in-process FakeMqttTcpServer. The per-operation cost includes the full publish → broker → subscriber round-trip (4× for QoS 1, 16× for QoS 2), so the Req/sec column represents the pair throughput (one send + one receive counted as 2 operations via OperationsPerInvoke = 1000 * 2).

MQTT 5.0 TCP

Via Mqtt5EndToEndTcpBenchmarks

BenchmarkDotNet v0.15.8, Linux Ubuntu 24.04.4 LTS (Noble Numbat)
Intel Core i9-9900K CPU 3.60GHz (Coffee Lake), 1 CPU, 8 logical and 8 physical cores
.NET SDK 10.0.103
  [Host]     : .NET 10.0.3 (10.0.3, 10.0.326.7603), X64 RyuJIT x86-64-v3
  Job-TMRHBV : .NET 10.0.3 (10.0.3, 10.0.326.7603), X64 RyuJIT x86-64-v3

InvocationCount=1  LaunchCount=10  RunStrategy=Monitoring
UnrollFactor=1  WarmupCount=10

Method	QoSLevel	PayloadSizeBytes	Mean	Error	StdDev	Req/sec
PublishAndReceiveMessages	AtMostOnce	10	3.243 μs	0.1560 μs	0.4599 μs	308,333.32
PublishAndReceiveMessages	AtMostOnce	1024	3.175 μs	0.3639 μs	1.0729 μs	314,927.69
PublishAndReceiveMessages	AtLeastOnce	10	3.810 μs	0.4279 μs	1.2615 μs	262,493.28
PublishAndReceiveMessages	AtLeastOnce	1024	3.866 μs	0.3386 μs	0.9984 μs	258,697.40
PublishAndReceiveMessages	ExactlyOnce	10	8.862 μs	0.6206 μs	1.8299 μs	112,846.37
PublishAndReceiveMessages	ExactlyOnce	1024	9.694 μs	0.6006 μs	1.7710 μs	103,161.13

32 KB payload benchmarks failed with OperationCanceledException (30-second CTS timeout). This is a known benchmark infrastructure limitation with large payloads on in-process fake servers, not a production regression. See the same behavior in the MQTT 3.1.1 8 KB benchmarks. Large-payload performance is I/O-bound and will be measured separately against a real EMQX broker.

MQTT 5.0 TLS (TCP+SSL)

Via Mqtt5TlsEndToEndTcpBenchmarks

Uses an in-process FakeMqttTlsTcpServer with a self-signed certificate. QoS 2 is excluded from the TLS benchmark as the 4-step handshake compounded with TLS overhead produces unreliable results under the 30-second CTS timeout in the current in-process test setup.

BenchmarkDotNet v0.15.8, Linux Ubuntu 24.04.4 LTS (Noble Numbat)
Intel Core i9-9900K CPU 3.60GHz (Coffee Lake), 1 CPU, 8 logical and 8 physical cores
.NET SDK 10.0.103
  [Host]     : .NET 10.0.3 (10.0.3, 10.0.326.7603), X64 RyuJIT x86-64-v3
  Job-TMRHBV : .NET 10.0.3 (10.0.3, 10.0.326.7603), X64 RyuJIT x86-64-v3

InvocationCount=1  LaunchCount=10  RunStrategy=Monitoring
UnrollFactor=1  WarmupCount=10

Method	QoSLevel	PayloadSizeBytes	Mean	Error	StdDev	Req/sec
PublishAndReceiveMessages	AtMostOnce	10	3.993 μs	0.2875 μs	0.8477 μs	250,464.78
PublishAndReceiveMessages	AtMostOnce	1024	4.521 μs	0.2667 μs	0.7864 μs	221,190.37
PublishAndReceiveMessages	AtLeastOnce	10	5.228 μs	0.3560 μs	1.0496 μs	191,265.99
PublishAndReceiveMessages	AtLeastOnce	1024	5.733 μs	0.4364 μs	1.2868 μs	174,415.56

Regression Analysis vs MQTT 3.1.1

Reference baseline from Mqtt311EndToEndTcpBenchmarks, same machine and runtime (Linux, .NET 10.0.3):

QoSLevel	PayloadSizeBytes	MQTT 3.1.1 (req/s)	MQTT 5.0 (req/s)	Delta
AtMostOnce	10	318,961	308,333	−3.3% ✅
AtMostOnce	1024	327,116	314,928	−3.7% ✅

Verdict: no throughput regression. The observed −3-4% difference for QoS 0 is within the measurement noise (MQTT 5.0 QoS0/10B 99.9% CI: ±4.8%). MQTT 5.0 adds slightly larger CONNECT/CONNACK frames and an expanded property set but the encoding/decoding overhead is negligible in the steady-state publish/subscribe path.

For QoS 1 and QoS 2, no Linux MQTT 3.1.1 reference run exists for comparison; the MQTT 3.1.1 Windows/.NET 8.0 results in Performance.md are not directly comparable due to hardware and runtime differences. The MQTT 5.0 QoS 1 results (~262k req/s for 10B) are consistent with the expected QoS overhead pattern.

Benchmark Design Notes

Benchmarks use an in-process FakeMqttTcpServer / FakeMqttTlsTcpServer, not a real broker. This enables git clone && dotnet run -c Release reproducibility without external dependencies.
The OperationsPerInvoke = PacketCount * 2 setting accounts for both the publish and receive sides of each message, so 1,000 messages yield 2,000 operations per invocation.
RunStrategy=Monitoring with LaunchCount=10 / WarmupCount=10 produces statistically robust results but is intentionally more expensive than a micro-benchmark run.
For real-broker throughput data see Performance.md.