|
Abstract : |
This dissertation examines scalability issues in the design of operating systems for largescale, shared-memory multiprocessors. In particular, the thesis focuses on structuring issues as they relate to memory management. From a set of simple, well-known queuing network formulas, we derive a set of properties that describe sufficient conditions for an operating system to scale. From these properties we first develop a set of guidelines for designing scalable systems, and then develop a new structuring philosophy for shared-memory multiprocessor operating systems, called Hierarchical Symmetric Multiprocessing (HSM). HSM manages the system resources in clusters, using tight coupling within a cluster, and loose coupling across clusters. Distributed systems principles are applied by distributing and replicating system services and data objects to increase locality, increase concurrency, and to avoid centralized bottlenecks, thus making the system scalable. However, tight coupling is used within a cluster, so the system performs well for local interactions. HSM maximizes locality which is key to good performance in large systems, and systems based on HSM can easily be adapted to different hardware configurations and architectures by changing the size of the clusters. Finally, HSM leads to a modular system composed from easy-to-design and hence efficient building blocks. Memory management is a particularly challenging service to implement within the HSM framework, because it must provide the applications with an integrated and coherent view of a single system, while distributing and replicating services in order to fully exploit the hardware potential. We describe in detail the implementation of an HSM structured memory management subsystem, and evaluate the performance of our implementation on Hector, a prototype scalable shared memory multiprocessor., |
