Skip to content

OpenVinc vs. ROS 2 benchmarks

WSL2 · Fast DDS · 2026-10-28–29
OpenVinc21/21qualified formal runsOrdinary push
ROS 2 · Data Sharing21/21qualified formal runsTuned profile
ROS 2 · Default11/21qualified formal runsrmw_fastrtps_cpp

In this same-host, cross-process test, OpenVinc had lower P99 latency than the default ROS 2 profile at sizes where all three runs of both profiles qualified. Its latency was in a similar range to ROS 2 with Data Sharing tuning.

At 1 KiB, 1 MiB and 8 MiB—the sizes where all three default-profile runs qualified—OpenVinc’s P99 was 49.5–74.5% lower. Against tuned Data Sharing, OpenVinc’s P99 was 1.6–15.8% higher from 1 to 64 MiB and 21.6% higher at 128 MiB; OpenVinc was lower at 1 KiB. The tuned profile had a lower P99 at six of the seven sizes. All of these differences remain visible below.

This compares observed latency and complete delivery under the recorded workload. ROS 2 RELIABLE, BEST_EFFORT and OpenVinc’s overwrite-history contract provide different delivery guarantees. Qualified-run counts do not make those guarantees equivalent.

P50 is median latency; P99 describes the tail. Each number below is the median of three per-run percentiles, with the full three-run range retained. Each run contains 64 warmup and 1,000 formal messages; warmup is excluded from latency statistics.

The first four sizes run at 100 Hz. The 32, 64 and 128 MiB cases run at 25, 10 and 5 Hz respectively. All three profiles use the same requested rate within a payload size.

P99 latency for OpenVinc, the ROS 2 default profile and Data Sharing; values are in the table below.
Logarithmic vertical scale; lower is better. Points are medians of three per-run percentiles; bars span their full min–max range, not confidence intervals. × marks a profile/size with fewer than 3/3 qualified runs; its measurements remain visible.

Download chart: SVG · PNG · PDF

Payload / rateOpenVincROS 2 defaultROS 2
Data Sharing
Vs. tuned¹
1 KiB100 Hz0.1053/3 qualified0.3223/3 qualified0.2333/3 qualified-54.8%
1 MiB100 Hz0.5693/3 qualified1.1263/3 qualified0.5073/3 qualified+12.1%
8 MiB100 Hz4.0743/3 qualified15.9943/3 qualified3.5193/3 qualified+15.8%
16 MiB100 Hz7.7463/3 qualified55.4730/3 qualified7.6213/3 qualified+1.6%
32 MiB25 Hz36.4493/3 qualified697.0830/3 qualified32.6243/3 qualified+11.7%
64 MiB10 Hz79.5313/3 qualified1000.0962/3 qualified73.9083/3 qualified+7.6%
128 MiB5 Hz178.7213/3 qualified2860.5490/3 qualified147.0143/3 qualified+21.6%

Milliseconds. Each P99 is the median of three per-run values, not a percentile of pooled samples. Qualification checks delivery, rate and other conditions; it does not make the delivery guarantees equivalent. ¹ Difference = OpenVinc ÷ Data Sharing − 1; a positive value means higher OpenVinc latency.

Full three-run P99 ranges
PayloadOpenVincROS 2 defaultROS 2 Data Sharing
1 KiB0.099–0.1160.309–0.4400.198–0.241
1 MiB0.484–0.6361.105–1.1300.495–0.543
8 MiB3.462–4.59114.120–16.6003.264–3.679
16 MiB7.098–7.96146.504–57.2386.935–8.023
32 MiB35.434–130.122624.079–881.05630.490–49.393
64 MiB76.883–81.317297.296–2280.86668.486–465.514
128 MiB175.338–188.5652593.246–3854.930139.923–153.464

Milliseconds; ranges are the smallest and largest per-run percentiles. Failed rounds remain in these statistics. Per-round maximum latency and qualification results are in the CSV and full archive.

At 32 MiB, OpenVinc’s three per-run P99 values span 35.434–130.122 ms, versus 30.490–49.393 ms for Data Sharing. At 64 MiB the pattern reverses: Data Sharing spans 68.486–465.514 ms, versus 76.883–81.317 ms for OpenVinc.

A single median cannot establish which system is more stable across every workload. The CSV and full report also retain the maximum individual latency from each run.

Profile Settings used here
OpenVinc Ordinary push with an independently owned local receive message. Sender loans are excluded from the main comparison.
ROS 2 default profile Jazzy / rmw_fastrtps_cpp, RELIABLE, KeepLast, middleware-default transports without a forced transport, and the RMW’s default Data Sharing OFF. Unbounded byte-sequence messages.
ROS 2 Data Sharing tuning BEST_EFFORT, explicit Data Sharing ON, per-size bounded messages, synchronous publication, adjusted history resources and SHM-only discovery/control transport.

“Default” means the specific baseline above, rather than every ROS 2 backend or deployment. The tuned profile does not establish the best possible ROS 2 performance either.

Both ROS profiles disable RMW application loans and intra-process routing. A SharedPtr callback owns the independent local receive message. Data Sharing shares DDS history; copies between the application and DDS may remain. This page makes no end-to-end zero-copy claim. Fast DDS documentation

Depth is 32, reduced to 30 at 128 MiB. OpenVinc’s actual Bus allocation is capped at 4 GiB. Equal depth does not imply equal total memory consumption.

The methodology was informed by Apex.AI’s performance_test project. This suite is independently implemented, without copying or linking that project’s code. The results here come from our own OpenVinc/ROS 2 workloads.

Item Recorded setup
Machine AMD Ryzen 7 5700X3D, 16 logical CPUs
OS WSL2 Ubuntu 24.04.5 on Windows 11
Build GCC 13.3, Release, native C++ APIs on both sides
ROS versions ROS 2 Jazzy; rclcpp 28.1.22, rmw_fastrtps_cpp 8.4.4, Fast DDS 2.14.7, Fast CDR 2.2.8
Processes One Publisher and one Subscriber in separate processes; OpenVinc also has a central supervisor
Timing boundary Publish API entry to receive-callback entry, after an independently owned local receive message is available; same-host CLOCK_MONOTONIC
Formal matrix 7 sizes × 3 profiles × 3 runs = 63 runs, executed individually with alternating profile order
OpenVinc revision a3a0843c5dc89f8b052c5f25afa66e1f171db99e, unmodified framework

These measurements were taken on 2026-10-28–29, using a different framework revision from the current v0.2.0. They are not a new measurement of that release. The archive retains versions, build configuration, commands, binary SHA256 values and a workload source snapshot.

Publishers reuse a preinitialized local byte vector and update its sequence for each message. Latency excludes that local buffer’s initial allocation. Receive checks validate length, sequence and edge markers, rather than a full payload hash.

Calibration selects rates before the formal configuration is frozen. Failed 100 Hz pressure pilots for the larger sizes remain in the archive. Qualification includes all expected IDs, valid markers and actual rate within ±5% of the target, among other checks. OpenVinc additionally checks collisions and overwrite/loss counters. Unavailable ROS internal counters remain null, not measured zeros.

  • This is a one-to-one, same-host WSL2 test with interactive desktop background load. No busy spin, CPU affinity, priority or global timer adjustments were used.
  • A 1,000-message run provides only about ten observations in its upper 1%. Three-run ranges expose variation but cannot establish long-duration or hard real-time guarantees.
  • These same-host WSL2 measurements do not cover networked hosts, other DDS vendors, fan-out, multiple subscribers or dedicated real-time systems.
  • Earlier UniquePtr callbacks introduced an extra deep copy. Those diagnostic runs remain under diagnostics; the entire primary matrix was rerun using SharedPtr.
  • The archive also includes one OpenVinc loan run per size. Its commit latency excludes preparation; full_path includes that preparation. These numbers are kept out of the ordinary push/publish plots and tables above.

The full report also contains CPU and RSS measurements. CPU sampling windows differ between roles, and RSS repeats shared pages across processes, so these values are not simply added to compare total CPU or physical RAM.

Dataset: 20261028-224514-735300. The complete ZIP contains the Chinese report, raw CSV, configurations and counters, calibration and failed attempts, provenance, source snapshot and license notices. It contains no executables or third-party runtimes.

The archived verification record reports identical raw-CSV P50/P90/P99/max recomputation for all 116 attempts. This page’s import independently recomputed all 63 formal runs and checked their three-run aggregates. Failed rounds were neither repaired nor removed.

To reproduce, obtain the suite base commit recorded in provenance/git.json, apply the archived source/ working-tree snapshot, then follow source/ros2/README.md. Use that preserved snapshot rather than substituting the benchmark repository’s current branch contents.