Granular enough for high-performance applications as scalable for storage systems.
Current storage systems are complex, monolithic, and challenging to upgrade once deployed. Designed and architected last decades, they do not benefit from the newest hardware like, e.g., persistent memory, high-performance networks, etc. Moreover, being developed for the millisecond scale, they do not support new applications like those envisioned by granular computing. Additionally, today's high-performance computing and Cloud's (HPCC) applications potentially exhibit heterogeneous data accesses (e.g., mixed bursts of writes and reads data IOs) given the adoption of Big Data and AI models on HPCC infrastructures. We are investigating and developing a new class of Storage Systems that can efficiently cope with these challenges and call them Granular Storage Systems (GSSs). GSSs can easily (and autonomously) adapt to new workloads (self-configurable) while dynamically changing their sub-components or granules. A storage granule is the smallest storage component that is replaceable fast enough (we target tens to hundreds of microseconds) to serve existing workloads better(!) while maintaining application integrity requirements. This research allows revisiting fundamental data tradeoffs and design principles for modern data storage systems at the granularity imposed by microsecond timescales.