AbstractPollingIoProcessor read buffer drain loop may block for substantial time.
This is because the QuickFIXj/MINA read filter chain eventually ends up in one of EventHandlingStrategy implementations (thread-per-session or shared), which in turn write messages into a fixed capacity blocking queue.

Reproduction - set up a NewOrderSingle generator and a simple Acceptor with delay of 20ms between subsequent ack ExecutionReport responses and a queue size of 1000. Pumping of 2000 NewOrderSingle messages into such Acceptor should be enough to see that execution reports are blocked at the Acceptor end until the sender stops sending new messages and the read queue of the Acceptor unblocks.

General idea of the fix - use watermarks-based queue instead of blocking queue. I'll do a github pull request shortly.