Measure the job, not the slogan
Your objectiveIdentify a comparison that can answer your actual decision.
Understand the idea.
Choose the task and required quality before choosing a metric. A throughput score can be accurate and still irrelevant to your workload. Record the inputs, configuration, concurrency, warm-up behaviour, failures and the time spent outside the measured component.
A worked example.
One service returns answers faster but fails more often on your document set. Comparing speed alone would hide a result that matters to the people using it. A successful completion needs a quality definition as well as a time measurement.
Try the reasoning.
Write a miniature benchmark plan with one task, one baseline, one success criterion and one failure criterion. Explain how you would prevent an easy input set from distorting the result.
Carry it into practice.
Keep the slow and failed cases. They often tell you more about operating constraints than the single fastest result.
Read the reference: TLB context: hardware performance claims ↗