lhotari commented on PR #24623: URL: https://github.com/apache/pulsar/pull/24623#issuecomment-3208984303
Interestingly, the results on my Dell XPS 2019 (i9-9980HK CPU @ 2.40GHz) are very different than on MBP M3 Max: ``` ┌────────────────────────────────────┬──────────┬──────────┬───────────┬──────────┬─────────────────┬────────┬────────┬──────────┬─────────────────────────┐ │ description │ produced │ consumed │ consumers │ bk_reads │ bk_read_entries │ hits │ misses │ restarts │ ts │ │ varchar │ int64 │ int64 │ int64 │ int64 │ int64 │ double │ double │ int64 │ timestamp │ ├────────────────────────────────────┼──────────┼──────────┼───────────┼──────────┼─────────────────┼────────┼────────┼──────────┼─────────────────────────┤ │ PR12258 │ 1509385 │ 15098443 │ 10 │ 18525 │ 1847641 │ 87.76 │ 12.24 │ 3 │ 2025-08-20 19:35:26.07 │ │ PR12258 │ 1693476 │ 16940269 │ 10 │ 13736 │ 1368859 │ 91.92 │ 8.08 │ 3 │ 2025-08-20 19:36:00.442 │ │ PR12258 │ 1723156 │ 17237529 │ 10 │ 12683 │ 1262572 │ 92.68 │ 7.32 │ 3 │ 2025-08-20 19:36:34.506 │ │ PR12258 │ 1719958 │ 17205041 │ 10 │ 13614 │ 1354546 │ 92.13 │ 7.87 │ 3 │ 2025-08-20 19:37:08.873 │ │ PR12258 │ 1644804 │ 16454401 │ 10 │ 11411 │ 1136525 │ 93.09 │ 6.91 │ 3 │ 2025-08-20 19:37:43.28 │ │ cacheEvictionByMarkDeletedPosition │ 1859918 │ 18607060 │ 10 │ 81565 │ 8129842 │ 56.31 │ 43.69 │ 3 │ 2025-08-20 19:38:27.66 │ │ cacheEvictionByMarkDeletedPosition │ 1814244 │ 18145240 │ 10 │ 81722 │ 8131713 │ 55.19 │ 44.81 │ 3 │ 2025-08-20 19:39:01.718 │ │ cacheEvictionByMarkDeletedPosition │ 1740813 │ 17412731 │ 10 │ 60897 │ 6042479 │ 65.3 │ 34.7 │ 3 │ 2025-08-20 19:39:35.845 │ │ cacheEvictionByMarkDeletedPosition │ 1664006 │ 16644406 │ 10 │ 38717 │ 3863385 │ 76.79 │ 23.21 │ 3 │ 2025-08-20 19:40:09.946 │ │ cacheEvictionByMarkDeletedPosition │ 1758310 │ 17593905 │ 10 │ 77274 │ 7686870 │ 56.31 │ 43.69 │ 3 │ 2025-08-20 19:40:43.946 │ │ PIP430 │ 1826902 │ 18272962 │ 10 │ 14760 │ 1471620 │ 91.95 │ 8.05 │ 3 │ 2025-08-20 19:41:28.689 │ │ PIP430 │ 1721983 │ 17226790 │ 10 │ 8366 │ 835071 │ 95.15 │ 4.85 │ 3 │ 2025-08-20 19:42:03.198 │ │ PIP430 │ 1711848 │ 17137830 │ 10 │ 1985 │ 197240 │ 98.85 │ 1.15 │ 3 │ 2025-08-20 19:42:37.248 │ │ PIP430 │ 1710176 │ 17111540 │ 10 │ 4288 │ 428004 │ 97.5 │ 2.5 │ 3 │ 2025-08-20 19:43:11.528 │ │ PIP430 │ 1690152 │ 16912942 │ 10 │ 1937 │ 192703 │ 98.86 │ 1.14 │ 3 │ 2025-08-20 19:43:45.748 │ │ PIP430disabled │ 1872724 │ 18734102 │ 10 │ 78763 │ 7842919 │ 58.14 │ 41.86 │ 3 │ 2025-08-20 19:44:30.079 │ │ PIP430disabled │ 1793152 │ 17940620 │ 10 │ 79898 │ 7936811 │ 55.76 │ 44.24 │ 3 │ 2025-08-20 19:45:04.486 │ │ PIP430disabled │ 1755762 │ 17563500 │ 10 │ 77663 │ 7726275 │ 56.01 │ 43.99 │ 3 │ 2025-08-20 19:45:38.471 │ │ PIP430disabled │ 1773824 │ 17743593 │ 10 │ 78873 │ 7846181 │ 55.78 │ 44.22 │ 3 │ 2025-08-20 19:46:12.815 │ │ PIP430disabled │ 1683544 │ 16839721 │ 10 │ 49009 │ 4875186 │ 71.05 │ 28.95 │ 3 │ 2025-08-20 19:46:47.135 │ ├────────────────────────────────────┴──────────┴──────────┴───────────┴──────────┴─────────────────┴────────┴────────┴──────────┴─────────────────────────┤ │ 20 rows 10 columns │ └──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘ ``` What seems to happen is that when CPU resources are scarce, differences in caching efficiency will accumulate, as the additional overhead from cache misses creates bottlenecks that propagate through the system, causing small efficiency gaps to manifest as significant performance differences under sustained load. The test case uses a PulsarMockBookKeeper, so storage operations themselves are fast. The test case doesn't add a penalty for reads at the moment. In reality, each read would occur more latency. This isn't a fully realistic test case, but it does show that PIP430 caching strategy is superior over others for the rolling restarts use case. One of the gaps of the test case is that it currently doesn't have a constraint on caching since there's no other active topics that would be filling the cache and causing it to overflow. A test where this happens would be useful since it could demonstrate the PIP430 strategy for handling cache overflowing. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: commits-unsubscr...@pulsar.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org