Token By Token

Episode 15 · Planned

Saturation, SLO and cost

Sweep concurrency and offered rate to find SLO-qualified goodput and attributable cost.

Hold input/output length fixed while sweeping concurrency. Separately hold the request mix fixed while sweeping offered arrival rate. Measure schedule lag and offered versus achieved rate so client overload is not mistaken for server capacity. Find the highest SLO-qualified goodput, not merely the highest raw throughput.

Evidence: predeclared TTFT/TPOT/end-to-end SLOs, warmup policy, request schedule, achieved rate, errors, client CPU/network load, latency distributions and cost per qualifying unit of work.

Read this experiment in the roadmap →