The split data cache in multiprocessor systems: an initial hit ratio analysis

Abstract

As current first level (L1) data caches are poorly and inefficiently managed, new approaches to achieve better performance in uniprocessor systems have been proposed. The L1 data cache management system is basically the same as it was three decades ago. New organizations have recently been proposed, where two multi-lateral caches are included in the first level in accordance with the data locality where they are stored. The processor simultaneously sends the same memory request to both caches located in L1. These caches work independently and have different organizations. The main objective is to minimize the average data access time. These new organizations will normally increase the hit ratio. Additionally, the chip area occupied by these caches-including the necessary management hardware-is smaller than in a conventional organization. As the proposed cache size is smaller, it can work faster and improve access time at this level. Several authors have studied different approaches around this idea in uniprocessors. In this work we have made extensions for shared memory multiprocessors and studied the advantages.