Hello, We are evaluating new hardware by reproducing real-life workload on real-life data which works fine on an existing server. We copied over the software used to ensure an apple-to-apple comparison, which is based on MariaDB 10.2.23 and uses primarily InnoDB tables, with a few Mroonga tables which do not seem involved in the problem. On the new hardware, we are seeing catastrophically bad performance, especially chain-deadlocks happening and being resolved but producing virtually no useful work at our usual parallelism level of 64 active connections. If we reduce parallelism to 2 active connections, queries start to succeed in any meaningful proportion. The database was freshly restored from a mysqldump (as we cannot interrupt the original database), could this have any effect on deadlocks ? The main tables involved in the deadlocking queries (job queues) are initially (in the test, and periodically in reality) empty, so it would seem surprising. What should I check to debug further ? Note that these deadlocks happen in production but they're negligible. See attachment for mariadb configuration file. Some more details: When we restore from mysqldump, we use the following extra options to speed things up: innodb_flush_log_at_trx_commit = 0 innodb_flush_method = nosync innodb_doublewrite = 0 sync_frm = 0 as during that period nothing of value can be lost (worst case we restart the restore from scratch). BTW, despite these settings, we are still noticing a lot of fsyncs. Is this expected ? Are we missing some other option ? Also, after the import and without any connection, MariaDB was still producing a non-trivial amount of activity on the machine: 5% CPU, read <1MB/s, write 10MB/s, 60 fsync/s. I could not identify what is causing these, where should I look ? How can I tell when it will stabilise back to idle ? When I interrupt the benchmark workload, I see a similar resource usage. I saw https://jira.mariadb.org/browse/MDEV-18698 And InnoDB seems to throttle its background activity: is there a way to tell InnoDB to perform its background tasks at maximum speed ? (the goal is that after the import, we can make a clean tarball that we extract when we want to launch the test case again) Regards, Julien