It seems specific problem only at the server. And it seems to be caused by multi sessions' load. (single thread load seems to safe)
I currently suspect "Hardware bug" of the server which like "breakable cache coherency" and cannot keep atomic operations.
The work area of malloc() seems to be broken by the multi-threads' load.
In the end, Percona-Server and XtraDB must be innocent.
It seems specific problem only at the server.
And it seems to be caused by multi sessions' load.
(single thread load seems to safe)
I currently suspect
"Hardware bug" of the server
which like "breakable cache coherency" and cannot keep atomic operations.
The work area of malloc() seems to be broken by the multi-threads' load.
In the end,
Percona-Server and XtraDB must be innocent.