<montywi> XL: I update the bb-fast-connect tree in git; This is after serg's review
<montywi> so I only need you to verify that things are same or better than normal 10.1 for connect tests and I can push
<XL> montywi: I see
<montywi> on my machine, having all threads in the thread cache gives a 5% speed increase; Having no thread cache was about the same
<montywi> this was when running 32 perl processes on the same machine to connect, do a query and disconnect
<montywi> doing this with a threaded benchmark should hopefully show a bigger speed increase

Pushed to bb-fast-connect tree for testing.

Results from a short benchmark. I tested 3 scenarios

1. standard sysbench OLTP readonly (1 connection per client thread that is reused)
2. sysbench OLTP readonly, reconnection for each transaction
3. simple workload: connect; select 1 from dual; disconnect

The last scenario shows advantages for the bb_fast_connect tree. The other however give better results with vanilla 10.1.

Benchmark was run on lizard2 (Intel, 32 cores, 64 threads). Huge thread pool was configured.

Pushed to 10.2.
Speed up for connections are up to 40% when there is a lot of connections threads and the thread pool is 16 or higher.

Re-opened for benchmarking the effects of the change