(SM)2-II: A Large-Scale Multiprocessor for Sparse Matrix Calculations

(SM)2-II is a large-scale parallel machine dedicated to scientific computation which includes sparse matrix calculations. In order to connect thousands of microprocessors and utilize a high degree of parallelism, the whole of the (SM)2-II system is consistently designed based on a simple computation model called the NC (NODE&CONNECTING-LINE) model. In this model, a target problem can be represented naturally with static concurrent processes (NODEs) and communication channels (CONNECTING-LINEs). A dedicated language NCC (Node oriented Concurrent C) has been designed based on this model, and a distributed processing system DIPROS supports a limited dynamic scheduling and allows each processor to manage multiple NODEs. A data-interchange mechanism called RSM (receiver selectable multicast) is proposed to support communication between NODEs. Each processor can multicast data with minimal conflict on the bus and minimal memory requirement. To connect a large number of processors, (SM)2-II is composed of multiple clusters, each consisting of multiple processors which are connected by the RSM. Intercluster communication is also based on the RSM principle. The connection structure of the cluster is organized so as to make the best use of the locality of communication involved in each problem. To demonstrate the efficiency of RSM, congestion on the system bus is analyzed. In addition, some cluster structures are evaluated to verify the performance of (SM)2-II.