pgpool performance issue

[liujinyang-highgo]

hi，
I am currently using Pgpool-II 4.6.2 and encountering a performance issue.
My architecture is Pgpool-II with 3 database nodes (1 primary, 2 standby), and I am running pressure tests via BenchmarkSQL.
The backend weight configured in Pgpool-II is set to 0:1:1.
When Pgpool-II is deployed on the primary database node, I observe that the tpmC is lower compared to directly connecting to the primary node. Has anyone encountered this issue before? What tuning methods are available?
Is it mandatory to deploy Pgpool-II on a standalone independent node?

[tatsuo-ishii]

Assuming there's enough memory and CPU is on the primary server, you may want to try to change the connection method from pgpool to PostgreSQL primary to Unix domain socket. i.e.:
backend_hostname0 = ''
This will reduce the network communication overhead, thus will contribute to performance if the networking is the bottle neck.
Of course this will not work after the primary PostgreSQL moves to another host where pgpool is not running though.

[liujinyang-highgo]

Modifying backend_hostname0 does bring some improvement, but the effect is not obvious.
At present, it appears that CPU is the real bottleneck. The combined overhead of the database and Pgpool is causing excessively high CPU pressure.
The purpose of using Pgpool is to reduce the load on the primary node. However, when Pgpool is deployed locally on the primary node, does this mean that the overhead introduced by Pgpool outweighs the benefits brought by read-write separation?

[tatsuo-ishii]

However, when Pgpool is deployed locally on the primary node, does this mean that the overhead introduced by Pgpool outweighs the benefits brought by read-write separation?

Yes, I think so. Pgpool-II needs to read and foward messages from/to clients and to/from PostgreSQL. That is really a CPU eater, especially in a benchmark software where small queries are dominant in the workload.