On Fri, 2021-10-01 at 14:22 +0530, Subhrajyoti Senapati wrote: > Was running a few sysbench tests in Postgres12. > Sysbench Test Config > oltp-readwrite-custom > Threads - 500 > Machine: 16 core 64G > > In PG server: > shared_buffers: 16GB > maintenance_work_memory: 16GB > checkpoint_timeout: 1h > max_wal_size: 10GB > wal_sync_method: 'open_sync' > effective_cache_size: 32GB > wal_buffers: -1 > wal_compression: on > > Got some qps dips. > > Changed fsync and full_page_rewrite to 'off'. > Which helped in avoiding these many long dips in performance: > > But fsync ensures DBMS consistency. So, instead of that can I change a few other configurations to avoid these inconsistent performance dips ?
Try playing with "bgwriter_flush_after", "checkpoint_flush_after" and "wal_writer_flush_after" and see if that influences the behavior.
