Enabling Scalable Performance for General Purpose Workloads on Shared Memory Multiprocessors

Taking a traditional operating system and making it scale well on shared memory multiprocessors is hard. Performance debugging a multiprocessor operating system is hard. Maintaining a scalable operating system is hard. We contend that designing a system from the ground up specifically for scalability can simplify these tasks considerably and allows for reasonably straightforward implementations that are scalable and maintainable. We have designed the K42 operating system from the start specifically for shared-memory multiprocessor scalability. Key to dealing with the complexities associated with scalability are (i) an object-oriented structure that maximizes locality, (ii) a Clustered Object infrastructure that supports distributed implementations in a straightforward way, and (iii) a hot-swapping infrastructure that allows one implementation to be exchanged for another at run-time as demands on objects change. In this paper, we describe K42’s design, analyze its scalability and performance using real workloads running on 24 processor commercial hardware, and show some of K42’s design tradeoffs as they relate to scalability.