Benchmark
Measured performance reports. Numbers are snapshots tied to the hardware and build noted in each report — not guarantees.
- Benchmark Report — Bare Metal — the reference numbers: end-to-end HTTP/MCP latencies and per-stage breakdowns on TencentOS bare metal, measured with the host stack and each microVM CPU/NUMA-pinned (
e2e/pin.sh). - Benchmark Report — Nested-virt — the same benchmark suite inside a GCP VM, with the measured why-it’s-slower analysis. Nested-virt runs are never pinned (the L1 vCPUs are already outer-scheduled).
- Benchmark Report — Bare Metal: microVM Reuse Capacity — the bare-metal capacity report: the ideal paced-lane test, the traffic-shaped test and the realistic AI agent model, each in pure-cold and mixed-turn shapes, with distinct whole-core allocations for loadgen / gateway / sandlot and the microVM pool. The physical-core mapping is the recommendation; the logical-core (SMT) mapping is a control (it halves the per-VM rate and adds no host capacity). Results are reported at three latency tiers (1.05x / 1.5x / 2x of the no-load baseline) with the absolute millisecond budgets. Raw logs are in
baremetal-load-raw/beside the report. - Benchmark Report — Nested-virt: microVM Reuse Capacity — three models on the same two-microVM host: the ideal paced-lane test (one lane per microVM, continuously fed), the traffic-shaped test (a full agent population queued per lane, one in-flight request per VM), and the realistic AI agent business capacity (0.9-1.5 s think time, 1-3 requests per turn, warm follow-ups inside the 1 ms sticky window, busy refusals retried like a real agent). Reports per-microVM and host cold reuses/s for each, the 1x/2x business capacity, and why a small nested host understates a production pool or fleet.
- Reproducing the Benchmarks — two ways to run the benchmark scripts yourself: the full developer build, or the prebuilt release package (no toolchain; benchmark scripts are bundled in the release tarball).