SHIP: A Storage System for Hybrid Interconnected Processors
Juan Camilo Vega · TSpace (University of Toronto) · 2020
Drivers for accessing storage are complex. In addition to the complexity involved in using the solid-state drive protocols, navigating filesystems is a process requiring multiple storage accesses and data processing between each access. As a result, efforts to create storage drivers for non-CPU processors (FPGAs and GPUs) have either failed, require too many resources/time, or remove some of the functionality expected by storage users (such as removing the filesystem). In this thesis, we explore the creation of a wrapper encapsulating a solid-state drive that performs all of the filesystem operations, and presents a much simpler network-based interface, simple enough for FPGAs and GPUs to use efficiently. Data transfers are performed via the network-based RDMA protocol, for which drivers exist for CPUs, GPUs, and FPGAs. ASIC accelerators for RDMA are also available. This system can sustain a throughput of 294 MBps, offers better scalability and is better adapted for cloud architectures, compared to current storage solutions