Start here. This is the direct spoken answer to practice first.
Why this question matters
Parallel code adds scheduling, partitioning, coordination, and merge cost. A faster inner loop can still make the service slower by consuming shared resources and worsening tail latency.
I would compare the parallel implementation with the original sequential version using representative data and the same correctness checks. I would measure elapsed time, throughput, p95 and p99 latency, CPU utilization, allocation, and error rate across several degrees of parallelism. The test needs enough work to include scheduling and merge overhead, not only the expensive function in isolation. I would keep the change only if it improves the metric the product cares about without violating resource or latency limits.