dPool : a distributed data structure for factored operating systems

Abstract

Future computer architectures will likely exhibit increased parallelism through the addition of more processor cores. Architectural trends such as exponentially increasing parallelism and the possible lack of scalable shared memory motivate the reevaluation of operating system design. This thesis work takes place in the context of Factored Operating Systems which leverage distributed system ideas to increase the scalability of multicore processor operating systems. fos, a Factored Operating System, explores a new design point for operating systems where traditional low-level operating system services are fine-grain parallelized while internally only using explicit message passing for communication. fos factors an operating system first by system service and then further parallelizes inside of the system service by splitting the service into a fleet of server processes which communicate via messaging. Constructing parallel low-level operating system services which only internally use messaging is challenging because shared resources must be partitioned across servers and the services must provide scalable performance when met with uneven demand. To ease the construction of parallel fos system services, this thesis develops the dPool distributed data structure. The dPool data structure provides concurrent access to an unordered collection of elements by server processes within a fos fleet. Internal to a single dPool instance, all communication between different portions of a dPool is done via messaging. This thesis uses the dPool data structure within the parallel fos Physical Memory Allocation fleet and demonstrates that it is possible to use a dPool to manage shared state in a factored operating system's physical page allocator. This thesis begins by presenting the design of the prototype fos operating system. In the context of fos system service fleets, this thesis describes the dPool data structure, its design, different implementations, and interfaces. The dPool data structure is shown to achieve scalability across even and uneven micro-benchmark workloads. This thesis shows that common parallel and distributed programming techniques apply to the creation of dPool and that background threads within a dPool can increase performance. Finally, this thesis evaluates different dPool implementations and demonstrates that intelligently pushing elements between dPool parts can increase scalability.