You're reading the documentation for a development version. For the latest released version, please have a look at Kilted.

Benchmarking

Benchmarks for the Zenbedded transport - Zenoh vs DDS (micro-ROS), Tier 1 vs Tier 2, and max control frequency across data sizes.

Default test setup: ESP32-S3, WiFi, Zenoh client mode, Best Effort QoS, ROS 2 Lyrical. All values TBD until measured.

Data sizes are expressed in float64 values, state to command at a 2:1 ratio. Tier 2 schemas carry exactly the fields they define.

Tier 1 vs micro-ROS (Zenoh vs DDS)

Same CDR messages (JointState, JointCommand), different wire protocol.

Metric	micro-ROS (DDS)	zenbedded Tier 1 (Zenoh)
MCU to ROS 2 latency (ms)	TBD	TBD
ROS 2 to MCU latency (ms)	TBD	TBD
Round-trip latency (ms)	TBD	TBD
Jitter p99-p50 (us)	TBD	TBD
Max control frequency (Hz)	TBD	TBD

Tier 1 vs Tier 2 (serialization)

Same user data, different wire cost. Tier 1 sends standard CDR messages; Tier 2 sends #pragma pack(push, 1) flat structs for the same values. Grid below matches the max frequency table.

State doubles	Tier 1 payload (bytes)	Tier 2 payload (bytes)	Tier 1 serialize (ns/op)	Tier 2 serialize (ns/op)
4	TBD	32	TBD	TBD
8	TBD	64	TBD	TBD
16	TBD	128	TBD	TBD
32	TBD	256	TBD	TBD
64	TBD	512	TBD	TBD

Both tiers carry the same Zenoh transport framing on top of the payload, so the framing cancels out of this comparison.

Max frequency vs data size

Publish at 50, 100, 200, 500, 1000 Hz for 10 s each. Max freq = highest rate with ≥ 99.9 % delivery. 1000+ samples per measurement, 100 warmup.

State doubles	Command doubles	Tier 1 max freq (Hz)	Tier 2 max freq (Hz)
4	2	TBD	TBD
8	4	TBD	TBD
16	8	TBD	TBD
32	16	TBD	TBD
64	32	TBD	TBD

Notes

Timestamps: k_uptime_get_32() on firmware (ms), rclpy.now() on host. One-way latencies assume synced clocks (NTP/PTP); round-trip needs no sync. ESP32-S3 is the primary target; native_sim results are supplementary.