Connection tests measure how many client connections a broker can accept and how connect latency behaves as the number of concurrently connected clients grows. Each scenario connects clients in batches and records, per interval, how many connections completed and their average latency.

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.

500K Connections

Scenario: 500K Connections - 500K-Connections-5000-rate.json for the 2026-11-02 runs, -10000-rate.json before them
Date: 2026-10-16
Client: AWS c5n.4xlarge, 29 bound source addresses, xmq_scn 0.9.20, no CPU affinity

QoS 1, 16B payload, MQTT 5; connect rate 10000/s before 2026-11-02 and 5000/s for the runs added then.

Broker ▲	Version	CPU Load	Max RAM	Avg RAM	Median RAM	Avg Connections	Avg Latency	Median	p99	p99.9
EMQX	6.3.1	909% mean / 1154% peak	12.88 Gb	10.1 Gb	10.7 Gb	500,000	13.2ms	13.2ms	—	—
FlashMQ	FlashMQ-1.27.2	34% mean / 91% peak	2.19 Gb	1.42 Gb	1.46 Gb	500,000	360us	342us	—	—
Mosquitto	2.1.2	29% mean / 39% peak	2.41 Gb	1.85 Gb	1.93 Gb	500,000	1.6ms	425us	—	—
XMQ	XMQ-1.1.1	50% mean / 95% peak	1.46 Gb	981 Mb	1015 Mb	500,000	323us	322us	—	—

EMQX

FlashMQ

Mosquitto

XMQ


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.

MQTT Performance Tests: Concurrent Connections

Connections

500K Connections

Connection latency, depending on the number of connections

Broker memory (RSS), depending on the number of connections