Exploiting MISD Performance Opportunities in Multi-core Systems

A number of important system services, particularly network system services, need strong scaling on multi-core systems, where a fixed workload is executed more quickly with increased core counts. Unfortunately, modern multiple-instruction/multipledata (MIMD) approaches to multi-core OS design cannot exploit the fine-grained parallelism needed to provide such scaling. In this paper, we propose a replicated work approach to parallelizing network system services for multi-core systems based on a multiple-instruction/single-data (MISD) execution model. In particular, we discuss the advantages of this approach for scalable network system services and we compare past methods for addressing the challenges this approach presents. We also present preliminary results that examine the viability of the basic approach and the software abstractions needed to support it.