| Flash storage has rapidly grown over the last few years into a redoubtable competitor to disk drives, due to its lower power consumption, lower latencies and lower cost per IOPs. Along with its attractive characteristics, flash also presents challenges, such as limited write endurance and the inability to overwrite a block in place. To address these issues, a software called the flash translation layer (FTL) maps logical blocks to physical locations on flash. Externally, the FTL presents a normal hard disk interface; internally, it implements address translation, garbage collection and wear-leveling algorithms which spread writes uniformly across the device.;In this thesis we provide a detailed investigation of flash device characteristics and internal logic, drawing valuable insight regarding their impact on higher-level properties such as whole-device endurance and throughput efficiency. Based on the derived characteristics, we model device endurance as a function of both the parameters of the chip itself, and the details of the internal algorithms used. We also construct analytic and black-box models to predict performance of solid-state drives under real workloads, and validate them by measurements, both in simulation and on real devices.;In addition, we investigate new design choices, specifically the integration of active computation capability in SSDs. We propose a novel approach, Active Flash, to migrate data analysis in scientific computing to the location of the data, the flash device itself; this can significantly reduce bandwidth, energy and time costs in high-performance computing clusters. We explore energy and performance tradeoffs in moving computation from host to storage, demonstrate the ability of embedded controllers to perform data analysis and reduction tasks at acceptable speeds, present a simulation study of scheduling policies, and implement an Active Flash prototype. These results show the viability of the Active Flash model, and its capability to potentially have a transformative impact on scientific data analysis. |
