Collaborative Research: CNS Core: Medium: Reconfigurable Kernel Datapaths with Adaptive Optimizations

Operating systems are a key part of every computer system. They sit between applications and hardware to provide security isolation as well as the sharing of physical resources such as the CPU, memory, and disk. Since operating systems run security critical code, they are hard to change safely, yet the policies they use for resource management can have a large impact on application and user-perceived performance. This project aims to develop an easily deployable architecture for lightweight data collection in the kernel, and then to use machine learning on that data to discover and deploy better kernel policies without compromising kernel security. The key idea is called reconfigurable kernel datapaths. Many parts of the operating system kernel process requests - to send network packets to the Internet, to store or look up file data, and to start and stop applications. Each of these involves a number of policy questions that could potentially benefit from machine learning. For example, the file system is often used in a repeated way, and by learning those patterns, it can potentially anticipate application behavior to bring data from disk into memory before it is requested. Operating systems do this to some extent already, but the policies they use are only weakly informed by data and can be counterproductive on certain access patterns. Reconfigurable kernel datapaths aims to address this, by providing a standardized architecture for instrumenting kernel datapaths, gathering data on the effectiveness of kernel policies, using machine learning to discover better policies, and finally deploying those policies using the same reconfigurable kernel datapaths used to gather data in the first place. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.