In the realm of array processing with its large data structures, identifying
no longer needed objects on the heap for reuse or timely release is crucial.
The only viable approach when using implicit memory management thus is garbage
collection via reference counting.  However, maintaining reference counts in a
highly concurrent context poses a mayor challenge.  Concurrent accesses to
shared data-structures require the corresponding, shared reference counter to
be updated as well.  To perform these operations safely, the reference counter
needs either to be protected by a lock or copied individually for each
concurrent access.  The former approach quickly leads to lock contention,
whereas the latter approach incurs additional costs for maintaining those
copies.

We propose to implement reference counting by asynchronous message passing
instead.  Modern many-core architectures like the experimental Microgrid or
Intel's SCC allow for low-latency high-speed ordered communication between
processor cores.  This makes it feasible to assign memory locations to
dedicated cores that manage the corresponding reference counts.  Reference
counting operations can then be issued concurrently by any core by means of
asynchronous inter-core communication.  These are automatically sequentialised
by the hardware and then processed in order by the receiving core.  We expect
the additional latencies from inter-core communications and the reference
counting operations themselves to have little impact on the overall runtime
due to the asynchronous nature of our approach.