TurboMqtt Performance

February 22, 2026 · View on GitHub

v1.0 Benchmark Update

Updated for v1.0 on Linux Ubuntu 24.04, .NET 10.0, BenchmarkDotNet v0.15.8 (i9-9900K). Original Windows/.NET 8 results are archived below for historical reference.

Current Benchmarks (v1.0)

MQTT 3.1.1 Results

For detailed results including raw per-launch data, see performance/mqtt311-benchmarks.md.

QoSLevel	PayloadSizeBytes	Mean	StdDev	Req/sec
AtMostOnce	10 B	3.313 μs	0.446 μs	301,798
AtMostOnce	1 KB	3.220 μs	1.276 μs	310,516
AtMostOnce	2 KB	3.731 μs	1.424 μs	268,033
AtLeastOnce	10 B	3.878 μs	1.875 μs	257,854
AtLeastOnce	1 KB	3.904 μs	0.996 μs	256,154
AtLeastOnce	2 KB	3.544 μs	1.084 μs	282,199
ExactlyOnce	10 B	8.488 μs	1.942 μs	117,816
ExactlyOnce	1 KB	10.964 μs	4.984 μs	91,209
ExactlyOnce	2 KB	9.971 μs	3.272 μs	100,289
ExactlyOnce	8 KB	13.513 μs	3.425 μs	74,005

MQTT 5.0 Results

For detailed results, see performance/mqtt5-benchmarks.md.

QoSLevel	PayloadSizeBytes	Mean	StdDev	Req/sec
AtMostOnce	10 B	3.243 μs	0.460 μs	308,333
AtMostOnce	1 KB	3.175 μs	1.073 μs	314,928
AtLeastOnce	10 B	3.810 μs	1.262 μs	262,493
AtLeastOnce	1 KB	3.866 μs	0.998 μs	258,697
ExactlyOnce	10 B	8.862 μs	1.830 μs	112,846
ExactlyOnce	1 KB	9.694 μs	1.771 μs	103,161

MQTT 5.0 vs 3.1.1 Comparison

QoSLevel	PayloadSizeBytes	MQTT 3.1.1 (req/s)	MQTT 5.0 (req/s)	Delta
AtMostOnce	10 B	301,798	308,333	+2.2% ✅
AtMostOnce	1 KB	310,516	314,928	+1.4% ✅
AtLeastOnce	10 B	257,854	262,493	+1.8% ✅
AtLeastOnce	1 KB	256,154	258,697	+1.0% ✅

Verdict: no throughput regression. MQTT 5.0 shows slight performance improvements despite larger CONNECT/CONNACK frames and expanded property sets. The encoding/decoding overhead is negligible in the steady-state publish/subscribe path.

Benchmark Reproduction

To reproduce these benchmarks locally:

# Run all benchmarks
dotnet run -c Release --project benchmarks/TurboMqtt.Benchmarks

# Run only MQTT 3.1.1 end-to-end benchmarks
dotnet run -c Release --project benchmarks/TurboMqtt.Benchmarks -- --filter '*Mqtt311EndToEndTcp*'

# Run only MQTT 5.0 end-to-end benchmarks
dotnet run -c Release --project benchmarks/TurboMqtt.Benchmarks -- --filter '*Mqtt5EndToEndTcp*'

See benchmarks/README.md for detailed reproduction instructions.

We designed the benchmark to include everything in order to make it git clone + dotnet run -c Release runnable right out of the box. That's the standard for best developer experience, but that means making some comrpomises on accuracy.

Why the `10b` message size?

We stuck with a relatively low message size because doing anything larger is mostly a matter of scaling Socket buffer sizes, and when we perform end-to-end benchmarking with real brokers like EMQX larger message sizes create memory pressure + availability problems for the broker itself. You can run these yourselves at larger sizes.

But the TL;DR; is: big messages are largely I/O bound problem - the whole purpose of TurboMqtt is make sure your publishing / receive message processing rates aren't CPU bound. Smaller message sizes make that easier to observe.

Data with Real Brokers

What sort of performance can you expect from real brokers?

TurboMqtt reading messages off of EMQX via MQTT 3.1.1

Using some of our application-specific stress tests, we've observed processing rates of ~70k msg/s (as fast as the stress testers could go against a single EMQX instance) running on QualityOfService.AtLeastOnce (QoS=1) - which is significantly faster than our benchmark data (see, we told you!)

Testing with larger packet sizes, such as 8kb packets, we've observed rates of around 35k msg/s, which translates to roughly 280 mb/s with a single client receiving messages. At those sizes you'll start to run into problems with your message broker long before TurboMqtt has any problems.