We present a new approach to implementing real-time transactions on memory-resident data on sharedmemory multiprocessors. This approach allows hard deadlines to be supported without undue overhead. In our approach, transactions are implemented by invoking wait-free library routines. Concurrency control is embedded within these routines, so no special support for data management is required of the kernel or from underlying server processes. These routines reduce the overhead involved in executing transactions by exploiting the way transactions are interleaved on priority-based real-time systems. We present evidence that shows that our approach often entails substantially less overhead than more conventional priority inheritance schemes.
James H. Anderson, Rohit Jain, Srikanth Ramamurthy