Worker Threads vs One Event Loop

Worker threads

Traffic

Presets

Request in flight Answered in time Missed the 100 ms deadline Dots are a small sample of requests. Numbers and timings are modeled.

Waiting for the first sampled request...

History (modeled): on-time throughput, event loop CPU, timeouts

On-time throughput (0 to 3500 req/s) Event loop CPU (0 to 100%) Timeouts (0 to 100%)

Modeled, not measured. Each request costs 5 ms of CPU in a worker (parse and match), then 3 trips through the event loop (to augmenters, to bidders, back to the worker). One loop message costs 80 µs with logging inside the loop (72 µs without), and every extra worker thread adds 2.5% to that cost (cache and lock contention). Augmenters take 8 to 16 ms and bidders 14 to 26 ms. The loop discards messages of requests that are already late, which is cheap.