Broadening the Use of GPU4FS

Having developed a completely GPU-side file system, we investigate the utility of our system to further parts of a larger computing system. In particular, we focus on three areas:

First, we investigate fast CPU-GPU communication channels. Even though the file system runs completely on the GPU, certain tasks like dynamic memory management require access to the page tables, which currently are only available to the CPU. We offer the established GPU-internal communication channels developed for GPU4FS for further uses like dynamic memory management and dynamic memory swapping between GPU VRAM and CPU DRAM guided by file system usage and requested via the file system communication channels.

Second, we enable the CPU to access the GPU-side file system. In certain CPU-intensive, parallel applications like 3D rendering or video processing, the file system latency is less relevant, but the application benefits from extra compute bandwidth. Here, we offload the file system accesses to the GPU. While latency increases due to the extra communication, the CPU usage of the file system decreases dramatically, and the application finishes faster.

Third, we extend the usability of the communication channels developed for GPU4FS. Dynamic work distribution inside a single and especially between different GPUs is complicated by the lack of a preemptive scheduler and difficult communication. We use the communication channels developed for GPU4FS to enable dynamic work stealing inside a single and between multiple GPUs, thus enabling new tasks to efficiently run on both single and multiple GPU systems.