Efficient local data movement in shared-memory multiprocessor systems

The DASH research project is addressing the general problem of achiev-ing high-performance network communication in large-scale distributed systems. The efficiency of moving a large amount of data between virtual address spaces (both user and kernel) on a single machine is a major com-ponent of this problem. Virtual memory (VM) remapping, as opposed to memory copying, is an attractive approach to moving data. However, remapping in shared-memory multiprocessors can be costly due to the problem of translation lookaside buffer (TLB) inconsistency. This paper describes the design of the DASH mechanism for moving data between virtual address spaces. This design integrates interprocess com-munication (IPC), virtual memory, and process scheduling mechanisms. By adopting a particular choice of IPC semantics based on a protected shared memory model, we are able to eliminate many of the overheads that would otherwise arise from VM remapping in shared-memory mul-tiprocessors. Put simply, we reduce the need for synchronous unmapping and, when it is necessary, we do it efficiently. 1