Hardware/software co-design for storage system

Abstract

In the past, storage systems served the simple purpose of data storage. However, in contemporary times, there is a trend towards designing storage systems that cater to diverse user requirements, particularly in cloud and data center environments. Despite this, the challenge of redesigning storage systems to meet user needs persists due to the intricate software stack and internal hardware structures they possess. This dissertation proposes a system that meets user requirements by concurrently designing hardware and software based on a comprehensive understanding of storage systems. The proposed solutions address three primary user requirements. Firstly, in a cloud environment with multiple containers, designing a storage system as if each container operates independently requires careful consideration. Chapter 2 tackles this by isolating input/output resources per container to ensure consistent I/O performance for each. Secondly, in a log-structured merge key-value store execution environment, ensuring uniform performance for all read latencies is imperative. Chapter 3 designs a storage system that guarantees a specific level of latency for read requests. Lastly, in an environment offering deep learning services for large-scale graphs, there is a demand to accelerate application services. Chapter 4 designs a dedicated computational storage for graph deep learning services.