Recently, a scalable Local Area MultiProcessor(LAMP) system has been proposed for distributed shared memory multiprocessors. SCI is an IEEE standard which defines an interface standard for very high performance multiprocessor systems. Remote shared memory access latency plays an important role in the performance of the LAMP system. A cache for a remote shared memory can be used to get over the performance reduction due to the remote shared memory access.
In this thesis, we design a software cache structure and remote memory access scheme to reduce remote shared memory access latency in the LAMP system using PCI-SCI bridge board which was developed in Center for Artificial Intelligence Research, KAIST. The proposed software cache includes coherence protocol using mail, dmove64, and write_sb4 transactions only. The data structures for the cache lines of the proposed software cache utilizes the memory-mapped features of the PCI-SCI bridge board to get a fine performance.
Simulation results show that the system using the proposed software cache can reduce remote shared memory access latency by 8%~50% of that of the system using no cache. The proposed software cache can get a good latency reduction on remote shared memory access if the work is partitioned carefully to reduce contention among processors and the data are aligned with cache line size. Especially, remote shared memory write latency is sensitive to the contention and the data alignment.