Report from Digitalocean managed Kubernetes cluster.
Cluster size: 3 Nodes CPU Optimized droplet - 8 vCPU 16 GB RAM
Estimated droplet cost's for cluster: $480/month
Mainflux services scaled to 3 instances:
- Mqtt adapter
- Authn
- Things
- Redis-things
- Envoy
- Nats
Publisher pool: 1
Subscriber pool: 1000
Test durition: 5min
RPS: 1
QoS Level: 2
In Fan-out test 1000 clients subscribe to same Mainflux channel and one publisher send one message every second over that channel to them. Both publisher and subcribers are connected with QoS 2, which is the safest and most demanding quality of service level that guarantees that each message is received exactly once by the intended recipients by using a four-part handshake.
No message loss was detected.
NOTE: Graphs Time unit is microsecond.
We see total number of connections during test is constant
We see total number of consumed messages is exactly 1000 times higher than the sent ones
*Pub to Pubrec latency in microseconds (100k microseconds are 100 milliseconds) *
Useful facts:
- 95% of clients had latency up to 12ms
- Max latency was up to 29ms
*Pub to sub latency in microseconds (100k microseconds are 100 milliseconds) *
Useful facts:
- 95% of clients had latency up to 130ms
- Max latency was up to 207ms
Results metrics
Results metrics are generated by mzbench tool
Kubernetes CPU usage in whole Mainflux namespace during testing
Kubernetes Memory usage in whole Mainflux namespace during testing
Kubernetes CPU usage in MQTT statefulset during testing
Kubernetes Memory usage in MQTT statefulset during testing
Kubernetes Network receive/transmit bandwidth in MQTT statefulset during testing