Numbers that hold up.
Reproducible measurements for DENSE — DE:LIGHT's KVM nanoservice execution mode. Hardware, workload, and methodology are published alongside every result.
Reference voting workload, measured in a KVM virtual machine (May 2026, published June 2026). See How we measured below.
Warm dispatch latency — end-to-end
The round-trip time from the host writing a request to the dispatch ring, through the KVM guest actor running, to the host reading the result. It does not include HTTP parsing, JSON serialization, or network time.
DEMIX sandbox figures are not published yet; they have not been measured on production hardware.
- Result
- 200 µs warm dispatch, 50th percentile (P50). The 99th percentile was about 380 µs.
- Workload
- Reference voting workload: three actors (validate, vote, tally), each in its own DENSE KVM guest, with cross-actor messages. One validate, one cast, and one tally trigger per run.
- Host
- AMD Ryzen 9 5950X, 64 GB RAM, Linux 6.12, running DENSE inside a virtual machine (nested KVM).
- Method
- 10,000 warm-up iterations, then 100,000 measured iterations. The benchmark process was pinned to one CPU core with the performance frequency governor, and the VM pool was pre-filled so no run includes a cold start.
- Dates
- Measured May 2026; published June 2026.
- Not published
- DEMIX sandbox spawn time has not been measured on production hardware, so no figure is shown.
These figures come from a virtualized environment; on bare metal we expect them to be equal or better. Bare-metal numbers will be published when measured.
echo_counter — Borz / Go / Node
The echo_counter workload is a minimal stateful counter with a single increment operation. It exercises dispatch, state access, and receipt generation without workload-specific hotpaths. Results across the three supported runtimes are published under the same methodology.
| Runtime | Warm p50 | Warm p99 | Notes |
|---|---|---|---|
| Borz | — | — | Results pending publication |
| Go | — | — | Same harness — results pending |
| Node | — | — | Same harness — results pending |
Results will be published once the validation series completes
live-rig runs. Raw results.json and
the reproducer script will be linked here alongside the numbers.
How the numbers are produced
Hardware
DE:EPER rigs — EU-hosted bare metal. Full specification and BIOS configuration published in the reproducer repository.
Isolation
Each run is pinned to a dedicated actor; no resource sharing between workloads during a measurement window.
Receipt inclusion
HATP receipts are generated for every call. The receipt latency is included in the reported wall-clock time — we do not measure a receipts-off fast path.
Reproducibility
The workload, harness, and analysis scripts are committed to a public repository. Anyone can run the same workload on comparable hardware and produce the same numbers.
EU-sovereign execution with a proof.
DENSE runs entirely on EU infrastructure. Every call produces a HATP receipt — a compact, verifiable proof of execution that can be presented to an auditor, a regulator, or a downstream system. The benchmark numbers above include that cost. There is no receipts-off mode that makes us look faster.