Design of a smart non-volatile memory controller: Architecture modeling, systems analysis, parallel I/O processing and scheduling algorithms

State-of-the-art Solid State Disks (SSDs) and Non-Volatile Memory (NVM) systems have undergone severe technology shift and architectural changes in the last couple of years, and, in parallel, SSD internal architecture has dramatically changed; modern SSDs now employ multiple internal resources such as NVM chips and I/O buses in an attempt to achieve high internal parallelism in processing I/O requests. In addition, to reduce intrinsic NVM system management overheads, SSD firmware employs advanced memory control strategies such as finer-granular address mapping algorithms and concurrency methods. As a result of complex interactions among these different mechanisms, modern SSDs can be plagued by enormous performance variations depending on whether the underlying architectural complexities and NVM management overheads can be hidden or not.;Designing a smart NVM controller is key hiding the architectural complexities and reducing the internal firmware overheads. To this end, we first model a multi-plane and multi-die NVM architecture, which is highly reconfigurable and aware of intrinsic latency variation imposed by diverse state-of-the-art NVM systems. This NVM model has been implemented as a high fidelity open-source simulator, capable of capturing cycle-level interactions between the many components in an SSD, which can be used for various high-level and low-level NVM performance analyses. Based on this architecture model, we then explore twenty four different concurrency methods implemented in NVM controllers, geared toward exploiting both system-level and NVM-level parallelism. Further, we quantitatively analyze the challenges, faced by PCI Express-based (PCIe) SSDs in getting NVM closer to CPU and question popular assumptions and expectations regarding storage-class SSDs through an extensive experimental analysis.;Next, we present and discuss the significance of read performance degradations and write performance variations by performing comprehensive empirical experiments using a diverse set of commercial SSDs and propose two novel schedulers in order to address these read/write performance challenges that modern SSDs face: 1) Physical Address Queuing (PAQ) scheduler and 2) NVM garbage collection scheduling algorithm. PAQ is a novel I/O request scheduling method that avoids resource contention resultant from shared SSD resources. Our proposed PAQ significantly improves read performance by exposing the physical addresses of requests to the scheduler and selecting groups of operations that can be simultaneously executed without major resource conflict. In comparison, the novel garbage collection scheduler is an approach that removes garbage collection overheads of underlying flash firmware and provides stable write performance in SSDs during the I/O congestion periods. Our proposed garbage collection scheduler tries to secure free blocks and remove on-demand garbage collections from the critical path in advance or delay them to future idle periods, so that users do not experience garbage-collection-induced latencies during the I/O-intensive periods. Overall, this thesis (1) presents a simulation infrastructure to conduct SSD/NVM research, (2) characterizes both system-level and device-level challenges faced by state-of-the-art SSDs, (3) presents a set of novel storage optimizations including various concurrency methods and scheduling algorithms design, and (4) points out future research directions.