Toward reconfigurable kernel datapaths with learned optimizations

Abstract

Today's computing systems pay a heavy "OS tax", as kernel execution accounts for a significant amount of resource footprint. This is not least because today's kernels abound with hardcoded heuristics that are designed with unstated assumptions, which rarely generalize well for diversifying applications and device technologies. We propose the concept of reconfigurable kernel datapaths that enables kernels to self-optimize dynamically. In this architecture, optimizations are computed from empirical data using machine learning (ML), and they are integrated into the kernel in a safe and systematic manner via an in-kernel virtual machine. This virtual machine implements the reconfigurable match table (RMT) abstraction, where tables are installed into the kernel at points where performance-critical events occur, matches look up the current execution context, and actions encode context-specific optimizations computed by ML, which may further vary from application to application. Our envisioned architecture will support both offline and online learning algorithms, as well as varied kernel subsystems. An RMT verifier will check program well-formedness and model efficiency before admitting an RMT program to the kernel. An admitted program can be interpreted in bytecode or just-in-time compiled to optimize the kernel datapaths.