Nice graphs sujatha.sivakumar. So 8-16 threads the throughput is higher. Still suffering on latency, particularly insert.

https://mariadb.org/fest2020/ssd/ at 10:48 offset - talking about fsync (redo, but same applies to binlog), that each fsync can be on the same data sector. Aligning every binlog unit to a beginning of a new 4k (discoverable fstat - blksize) block on disk after a fsync acceptable/show gains? And/or piggy back on the io_uring (MDEV-24883) implementation to have the kernel processing both binlog and other fsyncs for a transaction at the same tiem.

On a benchmark which aims to isolate the transaction commits alone, and varying different group commit parameters (i.e., binlog_commit_wait_count, binlog_commit_wait_usec, innodb_flush_log_at_trx_commit, sync_binlog, and concurrent connection count), we have the following results:

Added a new benchmark result prototyping the use of an innodb table to serve as the binlog (with the normal binary log disabled). Tested against various modes of flushing the binary log. In the legend, b stands for sync_binlog, and i stands for innodb_flush_log_at_trx_commit.

Running the innodb table prototype benchmark on an SSD (rather than ramdisk and WITH_PMEM) shows that with low concurrency, the innodb binary log implementation is able to have higher performance than the current file implementation; however, with more concurrency, the methods switch. Will do further analysis into the time spent, along with more benchmarks comparing differing binlog group commit parameters.

As far as I understand, implementing MDEV-34705 would significantly reduce the impact of setting sync_binlog. It should also make sync_binlog=0 (which would effectively be an alias of innodb_flush_log_at_trx_commit=0) crash-safe in a non-DDL workload. Yes, you might lose some latest committed transactions, but the binlog would be in sync with the storage engine.

For any crash safety, DDL operations seem to require fdatasync() or fsync(), as long as a separate ddl_recovery.log file is being maintained.