Moving threads: A non-conventional approach for mapping computation to MP-SOC

We present a completely new kind of approach for mapping the computation of an application to MP-SOC architecture. Instead of moving data read and write requests, extremely lightweight threads are moved between the processor cores. As a consequence, all kinds of cache coherence problems and need for read reply messages are eliminated. Lamport's sequential consistency of shared memory multiprocessor systems and flexible and efficient handling of varying number of threads are achieved for free. In our architecture, the challenge of having efficient implementation of an application reduces to mapping the used data so that the need to move threads is balanced with respect to the bandwidth of the local memories/intercommunication network. We outline required architectural techniques, apply them to a multithreaded shared memory MP-SOC framework and consider programming, compiling and runtime memory as well as thread handling issues. Few locality-aware mappings for common problems are demonstrated.