See the engine
under load.
Documented simulator runs connect the bid response, VAST delivery and spend accounting. Read the setup alongside the numbers.
Measured October 27, 2026 · Apple M4 · 10 cores · 16 GB memory
Inspect the variation
between runs.
Each open-loop run targeted 20,000 requests per second. The simulator and server shared the same laptop. These are client-side latencies, including loopback TCP, HTTP and scheduling.
The first run recorded 47 transport errors that did not reach the server. The second and third runs recorded none. Every request that reached the handler received a bid.
- Run 115.495 ms
19,890.5 requests/s · 1.210 ms p50 · 47 transport errors
- Run 241.407 ms
19,998.2 requests/s · 1.246 ms p50 · 0 transport errors
- Run 32.121 ms
19,998.3 requests/s · 1.120 ms p50 · 0 transport errors
Read the setup with the result.
- Host
- Apple M4, 10 cores and 16 GB memory, running macOS 27.0.1. The desktop session and simulator shared the host.
- Build and storage
- Rust release build with fat LTO and one codegen unit. The server used SQLite, an event spool and the example configuration.
- Buying catalog
- Four approved creatives and two line items, using video auction pricing and fixed-price audio. No ARTF agents, deals or frequency caps were enabled in these load runs.
- Request mix
- Audio streams, web video, CTV pre-roll pods and CTV live mid-roll, with randomized floors and private-marketplace opportunities.
These results describe a development snapshot on one shared laptop. Hardware, traffic mix, catalog size and external agent calls affect deployment performance.
Follow the bid
through delivery.
The 2,000-auction simulator run fired win, billing, loss and VAST player events. It recorded no HTTP, validation, timeout or transport errors in that run.
Explore event accountingThe simulator generated delivery events and read the resulting aggregates back from the server.
What else was measured?
A separate 10-second closed-loop run with 64 workers recorded 700,719 requests and 70,067.3 requests/s, with 2.673 ms client p99 and no reported errors. It describes throughput on the shared test laptop. Across the three open-loop runs, the server handler mean ranged from 44.9 to 56.7 microseconds.
Where do these measurements come from?
The engine repository's BENCHMARKS.md records the machine, commands, request mix, simulator reports, server metrics and method notes dated October 27, 2026. This page reproduces those documented measurements.
Switch to a modern,
affordable DSP.
Buy programmatically or go direct, with agentic workflows and validation included.