There's just way too much "semaphore" contention. Too many global mutex locks blocking threads that really aren't stepping on each other. Too many global mutexes held way too long. Disk i/o done by a thread holding a mutex!
MySql AB claims their product scales. Their proscribed technique is "scale out". But their "scale out" scheme is really a workaround for the fact their product doesn't scale.
There's a reason folks pay big bucks for Oracle Enterprise. Meanwhile, us entrepreneurial types trying to compute on the cheap have to grin and bear it.