Abstract

Accelerator-integrated Storage I/O (AiSIO) refers to a class of system software architectures that elevate accelerators such as GPUs, inference accelerators, and fixed-function units to first-class participants in the storage stack, enabling them to participate directly in data movement or initiate I/O themselves, while preserving compatibility and interoperability with existing operating-system managed storage abstractions such as files and file systems.

Modern accelerators provide massive parallelism and high-bandwidth local memory, yet their ability to access storage remains constrained by CPU-centered I/O paths and opaque, proprietary driver stacks that are difficult to extend or integrate with operating system infrastructure. As I/O demand continues to grow, these constraints lead to persistent underutilization of both compute and storage hardware.

AiSIO addresses these limitations through a structured separation between host-resident control-plane responsibilities and high-bandwidth data-path execution. The operating system retains authority over device management, metadata handling, and policy enforcement, while accelerators participate directly in data-path execution using mechanisms such as peer-to-peer DMA and accelerator-accessible I/O queues. This separation allows accelerators to initiate and drive I/O without undermining file system semantics, safety, or interoperability.

Host Orchestrated Multipath I/O (HOMI) serves as a reference implementation of these principles. HOMI explores how OS-managed, user space managed, and device-initiated I/O paths can be orchestrated to coexist with shared access to the same NVMe controller, either through software-mediated multiplexing or hardware-assisted delegation. A host-resident control plane coordinates device initialization, queue provisioning, and file-extent metadata, enabling accelerators to access file-backed data without requiring modifications to existing file systems.

This work presents HOMI as an experimental platform for studying accelerator-integrated storage I/O in open and modifiable system software. Through its design and evaluation, it demonstrates the feasibility of cooperative multipath I/O architectures that improve hardware utilization while preserving the guarantees and abstractions provided by operating-system managed storage.