UC San Diego

Emerging non-volatile memory technologies such as phase change memory, spin-torque transfer memory, and the memristor, will provide many orders of magnitude decrease in latency compared to disk and flash memory, dramatic increases in bandwidth, and a byte-addressable interface similar to DRAM. These new memories will offer enormous performance gains and intuitive abstractions for storage, but fully realizing these benefits requires us to rid software of disk-centric optimizations, design decisions, and architectures that limit performance and ignore bottlenecks previously hidden by the poor performance of disk. The algorithms that storage and database systems use to enforce strong consistency guarantees are critical to performance, and current solutions are deeply tied to conventional disk technology. This dissertation addresses the problem of providing transactional support for fast, non-volatile memories that exploits their raw performance and makes programming easy. First, we present a prototype PCIe-based storage array that targets fast, non-volatile memories and provides hardware support for multi-part atomic write operations. Multi-part atomic writes atomically and durably commit groups of writes to storage. Unlike previous approaches for flash-based SSDs, multi- part atomic write support makes logging scalable and transparent, providing a strong foundation for flexible ACID transactions. Using multi-part atomic writes, existing transactions mechanisms such as ARIES-style write -ahead logging can be redesigned to make optimal use of these memories, providing up to 3.7x the performance of the baseline version of ARIES. Second, we address the problem of providing strong consistency guarantees for storage that is directly accessible via the processor's memory bus. We present NV-heaps, a persistent object store which provides a familiar programming interface and protects against application and system failures by avoiding familiar programmer errors as well as new errors that only arise with persistent objects. Compared to Berkeley DB and Stasis, two persistent object stores designed for disk, NV-heaps improves performance by 32x and 244x, respectively, for operations on a variety of persistent data structures. To further improve safety, we present programming language support for NV-heaps. We introduce a Java-like language that provides the features NV-heaps require, along with a new static dependent type system that enforces the invariants that make NV-heaps safe