A Study of High-Performance Communication Mechanism for Multicomputer Systems

Based on our analysis of the behavior of standard UNIX communication software, we propose a new communication mechanism which reduces software overhead. This mechanism is optimized to be used in business applications, such as those of online transaction process ing and da tabase managemen t . These applications are currently designed to operate i n mul t ip le concurrent processes , thus requir ing multiprocessing environments with virtual addressing. The proposed communication mechanism, named DLA (DLA: Dynamic Link Allocation), supports direct data transfer communication between processors without any intermediate data copy operations. The DLA mechanism adopts a synchronous non-blocking interface, which can be used directly by user processes. Since the DLA mechanism provides address translation for each communication port, sender may directly specify the virtual address of the receive r’s buffer area. With DLA mechanism, sender and receiver CPU usages are reduced to one-third of those for TCP/IP communication, while attaining 30% higher throughput.