Why does TPC-D use a geometric mean for the power test?
• Since TPC-D is a representation of a generalized workload, all queries should contribute equally to the power metric; • Different implementations find different queries more “difficult” than others; • The metric should not unduly reward the “over tuning” of a single query. Because a metric based on an arithmetic mean of query execution times would be dominated by the longest running queries, the geometric mean was chosen for the power metric. To assure that an “over-tuned” query does not skew the results, the specification imposes a three-order-of-magnitude limit on the difference between the longest and shortest queries (Clause 5.4). Finally, it should be noted that the throughput test uses an arithmetic mean. In the simplest case of using a single query stream, the weight of any particular query in determining the throughput metric is directly proportional to its execution time (this is less clear for a multiple-stream case because there are multiple queries running concurrently).