Today’s computing has entered the cloud and AI era, where application workloads are highly diverse and dynamic, and the underlying hardware is increasingly heterogeneous. Operating systems remain the foundation that manages hardware and upon which all applications depend. However, achieving best-possible performance for each workload instance on a particular hardware setting remains difficult for two reasons. First, OS structure can prevent applications from fully exploiting hardware capabilities. Second, OS internals rely on static decisions that cannot fit all combinations of workloads and hardware. 

 

The goal of my research is to enable optimal performance across all workload instances. To achieve this, I build instance-customizable operating systems—OSes that can be customized at runtime to hardware and workloads without breaking compatibility. In this talk, I present two ideas that advance this vision. First, I introduce file systems as processes, which customize OS structure to better deliver storage performance to applications. Next, I describe an approach that customizes OS internals: performance tuning decisions as programmable policies. This approach introduces principles for flexible, safe, and fast in-situ tuning of previously fixed, performance-critical constants in deployed kernels, enabling millisecond-scale updates and up to 50 times performance improvements. Together, these ideas illustrate how operating systems can be made customizable per instance to deliver best-possible performance across diverse workloads and hardware, shedding light on the next generation of computing systems where in-situ code generation becomes a systems mechanism.