Currently a prepared statement works only for the thread/connection that actually prepared it. If a heavily multithreaded application runs the same queries in many threads, this leads to many duplicates in the prepared statement cache.

Example: sysbench OLTP. This one prepares a total of (2 + 9*number_of_tables) queries. Due to the above limitation it has to do this per thread. With for example 10 tables it cannot spawn more than 178 threads before reaching the default limit of 16K prepared statements.