Models direct device-to-device traffic. Each of 50,000 publishers has one subscriber on its own topic and sends it two messages a second. At QoS 1, the test measures end-to-end delivery latency at 100,000 messages/second over 100,000 connections. Every broker runs the same workload.

XMQ 1.0.2 and FlashMQ 1.27.2 were re-measured on 2 November 2026 on the same c5n.4xlarge host (16 vCPU, 40 GB) as the other brokers here. Percentiles are shown where the run recorded them; a dash means its result file has no percentile section.

QoS 1, 16B payload, MQTT 5, 10 minutes per scenario unless stated, connect rate 10000/s.

Broker ▲	Version	Messages	Achieved rate	Avg latency	Median	p99	p99.9	CPU	Avg RAM	Median RAM	Max RAM
EMQX	6.3.1	59,660,674	99,434/s	694.1ms	655.2ms	—	—	1529% mean / 1636% peak	4.78 Gb	4.58 Gb	5.74 Gb
FlashMQ	FlashMQ-1.27.2	59,999,980	99,999/s	230us	232us	—	—	352% mean / 408% peak	897 Mb	896 Mb	899 Mb
Mosquitto	2.1.2	59,729,696	99,549/s	2.2s	2.2s	—	—	99% mean / 101% peak	601 Mb	601 Mb	606 Mb
XMQ	XMQ-1.1.1	59,999,978	99,999/s	222us	222us	—	—	395% mean / 486% peak	1000 Mb	1000 Mb	1000 Mb

EMQX

FlashMQ

Mosquitto

XMQ

Delivery latency over the test

Broker memory (RSS) over the test

Latency uses a logarithmic axis: the brokers differ by several orders of magnitude in this scenario, and a linear axis would flatten the faster ones onto the baseline. Hover the chart for per-interval values.


Source code on GitHub
Credits:	The XMQ web site is provided by Linotex.	The XMQ and SPTK Windows installers are created with Advanced Installer free Open-Source license.