Bertels et al. study the problem of memory allocation in an environment that includes both a general processor and an attached hardware accelerator. In the model studied, memory is shared between the two processors, with all memory accessible to either processor. Still, local memory can be accessed much cheaper than memory attached to the other processor, so it is desirable to allocate each object to the processor that will access it more frequently.
The problem is studied through several models, including naive, ideal (Delphic), and heuristically chosen allocations. The authors show that a heuristic allocator performs very well in many important cases, and they do a fine job of discussing common cases and showing why their solution is effective and efficient.
In their future work, the authors should include studies of more complex usage cases, multiple attached processors, and other parallel models.