Which method is used to measure transformer performance by analyzing the translation log?

Which method is used to measure transformer performance by analyzing the translation log?

• A. Add up the time from all transformer-related INFORM messages in the Translation Log.
• B. Compare the runtime of a run with writers disabled vs with transformers disabled.
• C. Use a stopwatch from the first transformer to the end.
• D. Ignore logs and rely on feature counts.

Answer: (A)

Measuring transformer performance from the Translation Log relies on the timing data the engine records during processing. When a translation runs, each transformer emits an INFORM message that includes how long it took to process the related feature (or batch). By adding up all those durations for transformer-related messages, you get the total amount of time spent by transformers during that run. This provides a direct, in-engine view of where time is being spent, helping you spot bottlenecks by identifying transformers that accumulate the most time. This approach is practical and accurate because it uses the engine’s own timing data without altering the workspace or needing separate benchmark runs. It’s more reliable than hand-timing with a stopwatch or estimating time from feature counts, since counts don’t reflect actual processing time. Just keep in mind that the times are per feature and per transformer, so for consistent comparisons you should use the same dataset and consider aggregating to get stable results.

Measuring transformer performance from the Translation Log relies on the timing data the engine records during processing. When a translation runs, each transformer emits an INFORM message that includes how long it took to process the related feature (or batch). By adding up all those durations for transformer-related messages, you get the total amount of time spent by transformers during that run. This provides a direct, in-engine view of where time is being spent, helping you spot bottlenecks by identifying transformers that accumulate the most time.

This approach is practical and accurate because it uses the engine’s own timing data without altering the workspace or needing separate benchmark runs. It’s more reliable than hand-timing with a stopwatch or estimating time from feature counts, since counts don’t reflect actual processing time. Just keep in mind that the times are per feature and per transformer, so for consistent comparisons you should use the same dataset and consider aggregating to get stable results.