Next: Peer-to-peer communication.
Up: Current Issues and Future
As noted, checkpointing and migrating processes at all calls to bsp_sync() can lead to unnecessarily poor performance in programs where large amounts of local state persist between supersteps. We can improve efficiency by having bsp_sync() normally send back communication requests only, and only occasionally send back the entire process state. We can also implement a lazy migration scheme that does not migrate a virtual process unless it is causing the system to slow down unnecessarily.