Use different queue implementation

[insanitybit]

The double locking queue is probably a lot slower than required. There's a crossbeam dequeue that will likely work better.