As technology improves, changes in hardware drive corresponding adaptations in software. This thesis examines the hardware-driven evolution of both applications and system software as they relate to the matter of data storage in modern computing systems.

The move from single-processor systems to ones with numerous CPU cores executing in parallel has motivated applications to make increasing use of multithreading. The resulting nondeterminism introduces new difficulties to the common technique of evaluating storage system performance by replaying traces of application execution. We present the ROOT technique and implementation of it, ARTC, to address this challenge and provide a trace replay system for multithreaded applications that is both reliable and accurate in its performance projections.

Storage hardware has also undergone major changes in recent years, with traditional hard-disk drives increasingly displaced by flash-based SSDs, and even more recently emerging nonvolatile memory technologies. This shift drives the need for new software to manage these new devices and provide useful storage features and functionality, such as file cloning and deduplication, in a manner well-suited to the characteristics of the new hardware. Here we present ANViL, a storage virtualization system that provides these features in a novel way developed expressly for flash storage.

The dramatic difference in the performance characteristics of emerging storage technologies relative to the much slower mechanical devices they are replacing, however, also shines a new and unflattering light on the performance of storage software. Much of this software dates from the era of the hard-disk drive, when CPU cycles were often considered essentially "free" in comparison to the long latencies of disk operations. This performance imbalance made it easy to do relatively expensive things in software, safe in the knowledge that their performance cost would be hidden by the much slower storage devices they managed. However, as the performance gap between CPUs and storage hardware narrows, the CPU execution performance of software in the storage stack becomes increasingly critical. For this problem we present DenseFS, a prototype filesystem with the explicit aim of minimizing its use of CPU cache resources in an effort to not only run efficiently itself, but also to reduce its impact on application performance.

These pieces exemplify how software evolution in response to changes in hardware occurs, but also how it differs as the hardware in question becomes increasingly well-established. Multicore CPUs have been commodity items for over a decade and are now nearly unavoidably ubiquitous; we examine a delayed, second-order effect of this change on a specialized area of storage software, as its more immediate effects have been studied since it was a younger technology. Flash has been widespread for some time, but is not yet so deeply ingrained in the hardware landscape; here we examine one part of the ecosystem of storage software that is still in the process of adapting to suit the new hardware. Finally, NVM technologies are just beginning to arrive; the major, first-order questions it raises, such as what an NVM-oriented filesystem might look like, are thus still being addressed. Taken together then, these three components illustrate different stages in the chronology of how software's hardware-driven evolution has occurred, and how we expect it is likely to continue.