CPU overhead for trans_commit_stmt uses almost 2X more CPU in MyRocks than InnoDB

[mdcallag]

I am looking at flamegraphs from sysbench with a cached database and 1 client. A common problem I see is that the percentage of samples that lock_tables (and its callees) accounts for is ~2X larger in MyRocks than in InnoDB with simple microbenchmarks.

The microbenchmark in this case is here.

The percentage of samples from trans_commit_stmt and its children is 21.01% for InnoDB vs 35.19% for MyRocks

Percentage of samples, from InnoDB

21.01%  trans_commit_stmt
  17.97%  MYSQL_BIN_LOG::commit
    17.00%  MYSQL_BIN_LOG::ordered_commit
       3.60%  MYSQL_BIN_LOG::flush_cache_to_file
       3.05%  MYSQL_BIN_LOG::process_commit_stage_queue
       8.57%  MYSQL_BIN_LOG::process_flush_stage_queue
   1.27%. MYSQL_BIN_LOG::prepare

And from MyRocks

35.19%  trans_commit_stmt
  27.19%  MYSQL_BIN_LOG::commit
    26.19%  MYSQL_BIN_LOG::ordered_commit
       2.56%  MYSQL_BIN_LOG::flush_cache_to_file
      10.69%  MYSQL_BIN_LOG::process_commit_stage_queue
        ... mostly RocksDB code
      10.44%  MYSQL_BIN_LOG::process_flush_stage_queue
        ... mostly RocksDB code
   3.59%  MYSQL_BIN_LOG::prepare
     3.51%  ha_prepare_low
      3.32%  myrocks::rocksdb_prepare
        ... more MyRocks and RocksDB code     
   2.20%  is_thd_db_read_only_by_name

Flamegraphs