+
+ // Three numbers giving rules of thumb for judging our own implementation:
+ //
+ // - Compressed speed is easy to compare to disk I/O, to see the relative importance
+ // - Uncompressed speed is easy to compare to intersection speeds and memory bandwidth
+ // (also very roughly comparable to the benchmark numbers in the TurboPFor README)
+ // - ns/element gives an absolute measure for plocate (e.g. if we can decompress at
+ // 1 ns/element, a 10k-element posting list goes by in 0.01 ms, which is way beyond
+ // instantaneous in practice).
+ fprintf(stderr, "%.1f MB/sec (compressed), %.1f MB/sec (uncompressed), %.1f ns/element\n", posting_list_bytes / own_sec / 1048576.0,
+ (total_elements * sizeof(uint32_t)) / own_sec / 1048576.0, 1e9 * own_sec / total_elements);