Experimental analysis of synchronization methods for CC-NUMA systems

Non-Uniform Memory Access (NUMA) systems generally have two kinds of interconnection networks: one is a local network that connects a local memory and several processors within a node, and the other is an inter-node network that connects nodes. Intra-node transactions are performed within a node using only a local network, and inter-node transactions work with more than one nodes using both intra- and inter-node networks. It is very trivial that it takes inter-node transactions a longer time to finish their job than intra-node transactions. It consequently means that transactions for synchronization, such as acquire or release transaction, generally need a long time to be performed because synchronization variables are mostly shared by caches. From this point of view, we have proposed a new synchronization method called Freeze and Melt. In this paper, we describe its features and compare the performance of various synchronization methods using program-driven simulation.