Lab 2 (Ultra Disk), scenario H (hot-tail inserts): why parallel redo was slower. Measured on node-4 from the DMV collector samples (redo requests, wait statistics and latch statistics, every 1 to 5 seconds), over the window from the end of each workload to the end of its drain. Six legs per mode. Collector: scripts/40_dmv_collector.sql. Queries: scripts/41_analysis_queries.sql. measure serial (6 legs) parallel (6 legs) ------------------------------------------------------------- ---------------- ------------------------- drain time (s) 41.7 to 51.8 60.8 to 71.5 CPU used by the main redo thread (s) 39.5 to 45.7 52.4 to 58.6 main redo thread CPU per MB redone (ms) 6.1 to 7.0 8.0 to 9.1 main redo thread busy (share of one core) 88 to 95% 82 to 88% CPU used by all parallel redo worker tasks (s) none 34.3 to 53.8 CPU used by the busiest worker task (s) none 11.7 to 12.8 workers waiting for work, PARALLEL_REDO_WORKER_WAIT_WORK (s) none 240 to 285 RECOVERY_MANAGER latch waits (count; total s) not present 479,000 to 546,000; 6.2 to 7.1 other parallel redo waits: transaction turn, drain worker, flow control, log cache, transaction list, dirty page table (total s) none 0.7 to 8.0 page read waits, PAGEIOLATCH (s) 0.2 to 0.5 1.6 to 29.3 Contrast, scenario I (random-key inserts), same lab: parallel: workers used 189 to 202 CPU-seconds and waited 145 to 147 s on page reads per leg. serial: 73 to 74 CPU-seconds, 45 to 57 s on page reads, in a 269 to 275 s drain. Scenario L (wide inserts): main redo thread 6.0 to 6.6 CPU-s (parallel) and 6.8 to 7.8 CPU-s (serial) for 1.9 GB, about 3 to 4 ms per MB. Not established: which code path in the main redo thread costs the extra CPU. CPU profiles were collected only for runs A to E. The RECOVERY_MANAGER latch waits, absent in serial mode, point at the hand-off between the main thread and the workers; they don't prove it.