Environment Setup:
three 8 core 16GB Mongos
five shards each with 8 core 16GB Mongod
Workload:
SYSBENCH
Database: YCSB; Collection: t_0;
Set {field0: 1} as shard key， perform pure read workload with _id
We were surprised to find that when setting taskExecutorPoolSize to 8, the SYSBENCH qps ~1300, but when setting taskExecutorPoolSize to 1, the performance improve to ~5400 qps. Then we found this jira, but this jira explains little, can someone explain why setting taskExecutorPoolSize will cause such a difference.
We analyze the Flame Graph( it’s too big to upload, so we just show the pic)
when setting taskExecutorPoolSize = 8, we found there is a lot of futex syscall.
when setting taskExecutorPoolSize = 1, there is less lock operation.
