Remote-to-local block transfers are both simple and worthwhile to support in hardware. By reducing the time required to perform such a transfer, the page replication and migration operations that are common in NUMA shared memory multiprocessors can be more effective in dynamically placing memory blocks close to the processors that are accessing them. We have shown how our example multiprocessor can be enhanced to efficiently support block transfers. Finally, we should note that block size is an important parameter which has a significant effect on the overall performance. This has been considered in [2].
Providing efficient block transfers is only the first step. More aggressive memory management policies should be developed to take advantage of the relatively inexpensive block transfers.