DSZOOM – Low Latency Software– Based Shared Memory

Software-implementations of shared memory are still far behind the performance of hardwarebased shared memory implementations and are not viable options for most fine-grain sharedmemory applications. The major source for their inefficiency comes from the cost of interruptbased asynchronous protocol processing, not from the actual network latency. As the raw hardware latency of inter-node communication decreases, the asynchronous overhead in the communication becomes more dominant. Elaborate schemes, involving dedicated hardware and/or dedicated protocol processors, have been suggested to cut the overhead. This paper describes how all the asynchronous overhead can be completely removed by instead running the entire coherence protocol in the requesting processor. This not only removes the asynchronous overhead, but also makes use of a processor that otherwise would stall. The technique is applicable to both page-based and fine-grain software shared memory. Our proof-of-concept implementation—DSZOOM-EMU—is a fine-grained software-based shared memory. It demonstrates a protocol-handling overhead below a microsecond for all the actions involved in a remote load operation, to be compared to the fastest implementation to date of around ten microseconds. The all-software protocol is implemented assuming only some basic low-level primitives in the cluster interconnect. Based on a remote atomic and simple remote put/get operations the requesting processor can assume the role of the directory agent, traditionally assumed by a remote protocol agent in the home node in other implementations. The implementation is thread-safe and allows all processors in a node to simultaneously perform remote operations.