問題文
The same model must be benchmarked on an on-premises cluster and on rented cloud instances, and the two numbers must be comparable. Which controls make them comparable?
選択肢
- Fix the total wall-clock duration of each run and compare the number of requests completed; two runs of equal length measure the same amount of opportunity, and the completed count is a direct comparison of the two environments' capability.
- Fix the request pattern, the batching and concurrency settings, the precision, and the serving stack version, and report the hardware alongside each number.
- Run both with a single request at a time to remove the effect of batching, so that the only remaining difference between the two numbers is the hardware they ran on.
- Use each environment's recommended default settings so both are at their best, a comparison of two configurations at their respective peaks being the fairest one available.