Today I committed a change to Npgsql which will improve connection pool performance. This change was motivated by Andrew's bug report where he noticed that a lot of threads were waiting to get a new connection from pool.
In order to keep consistence of the pool, Npgsql has to lock access to it. Andrew's problem appeared in a busy server where a lot of threads were trying to get a new connection from the pool. They had to wait in line. And obviously this isn't good.
The current implementation of Npgsql creates a big lock surrounding all the code needed to work with the pool and more! As Andrew noticed in his bug report, I/O operations were being done inside this lock which was contributing to more delays to get a connection from the pool.
So, to fix that, I rewrote connection pool logic to remove this big lock and break it down to smaller ones only when really needed. All the I/O operations were left out of the locks, this way, other threads waiting to get a new connection from the pool don't need to wait for those expensive operations to finish.
I made a lot of tests and could confirm that when I break the code inside the debugger, threads are spread throughout connection pool code as expected instead of waiting in line on the big lock.
As this change is somewhat critical to Npgsql usage, I'd like to ask you to download the code, compile it and give it a try and see if everything is working ok or even better than before. I expect busy servers to be able to increase their raw throughput because it will have to wait less to get connections from the pool.
As always, please, let me know if you have any problems and all feedback is very welcome!