Ah, yes. It's good that you ran a benchmark. However, if I read the code correctly, your version with a lock does an extra memcopy. When the work function is as

Ah, yes. It's good that you ran a benchmark. However, if I read the code correctly, your version with a lock does an extra memcopy. When the work function is as short as AES encoding of a block of memory, that memcopy is quite a big cost and the queue is going to be quite heavily contended.