National Science Foundation Expeditions in Computing: Learning Directed Operating System -- A Clean-Slate Paradigm for Operating Systems Design and Implementation

Operating Systems (OSes) enable multiple application programs to run simultaneously on a computing device. To manage a device’s hardware resources (such as CPUs, GPUs, memory, and networking) in the presence of multiple running programs, OSes have so far relied on manually crafted heuristic policies that make broad assumptions about the applications that are likely to run and the environment in which they operate. Recently, there have been shifts in computer hardware technology and usage such as new CPU types, heterogeneous accelerators, and novel applications with complex changing needs running on the cloud and emerging platforms such as robots, autonomous vehicles, and edge computing. Unfortunately, heuristic policies, lacking complex and rich reasoning, work poorly with these advancements and result in poor performance and inefficient use. This has led to degraded user experience and higher costs. Manually customizing heuristic policies to meet the needs brought to the fore by each advancement is time-consuming, expensive, and ultimately untenable given rapid innovation. The Learning Directed Operating System Expedition (LDOS) aims to usher in a principled and sustainable solution to these challenges. LDOS is a next-generation OS that offers: (1) intrinsic intelligence—where advanced machine learning (ML) makes resource management decisions that maximize performance and efficiency, and (2) auto-adaptation—where the OS adapts to different settings with minimal human intervention. LDOS seeks a transformative impact on society by improving performance and decreasing inefficiencies and costs associated with new technologies. It could enable the creation of innovative and affordable computing devices, such as consumer-grade robots that assist humans in their day-to-day activities akin to smartphones today. LDOS could significantly improve the energy efficiency of large-scale cloud computing and artificial intelligence (AI) infrastructure. LDOS can pave the way for smart cities and factories by enabling novel real-time edge computing applications. LDOS’s auto-adaptation enables future devices to gracefully cope with unexpected changes (e.g., robots deployed in more crowded environments) without requiring extensive re-engineering. The LDOS Expedition, involving and fundamentally advancing multiple disciplines in computer science, rethinks OS design with ML-driven resource management at its center. LDOS offers a new class of ML-based policies driven by rich run-time data and trained using diverse synthetic data, building on fundamental advances in generative AI and ML algorithms. To ensure LDOS meets a wide range of application and system-level needs, ML model training will leverage verified learning, a novel integration of ML with formal verification techniques. LDOS’s ML-centric OS interfaces and abstractions will enable easy integration and automatic adaptation of ML policies, low-overhead ML-based decisions, and security and manageability. The project involves close collaboration with industry partners to create the open-source LDOS implementation and demonstrate compelling use cases such as autonomous personal robots, efficient and dependable cloud computing, and real-time cellular access edge computing. In addition to building an LDOS, the Expedition will leverage the popularity of ML to reboot excitement in computer systems and create a new curriculum around the interplay of computer systems and ML. The project’s initiatives are designed to cultivate leadership in ML and computer systems by offering tailored programs at different educational levels from middle school to higher education. This ensures participants are well-prepared for ML and computer systems technology and research careers, benefiting hundreds to thousands of students each year. 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.