Library support for historical and persistent data structures in non-volatile memories

Abstract

In the context of emerging non-volatile memory (NVM) where data structures can persist in-memory and are accessed through CPU loads and stores, we study how to efficiently manage data evolution. This is an extensively applied problem in both the scientific and business domains and is rapidly becoming an important component for a wider range of applications. We argue that the best way to achieve a smoother transition to the new programming model is to design a solution that is non-intrusive and generic i.e. not bound to a specific data model. We propose a novel library-level approach where the user can manage historical data directly from programming language code. This is achieved with a combination of two software layers: REWIND and VARIANT. At the bottom, lies REWIND (REcovery Write-Ahead System for In- Memory Non-Volatile Data Structures) which handles the low level specifics of NVM by dealing with write-ordering problems that arise in such context and allows recoverability of arbitrary data structures. Then, VARIANT (Versioning ARbItrary dAta structures in Non-volatile memory for Time-travel) focuses on versioning and time travel (moving between versions). We adopt a logging approach and we tightly integrate both systems for best performance by utilizing a common physical log of memory operations. With REWIND, we propose a novel recoverable log structure that permits atomic and durable appends and removals of log records. This is the keystone for building recoverable systems on top of NVM. Because latencies in recent NVM technologies such as Phase-change memory (PCM) are asymmetric, we propose novel techniques for reducing the write pressure of the recoverable log as well as mitigating the effect of synchronization control primitives such as memory fences (enhanced for NVM), i.e. barriers that enforce ordering and persistence to preceding instructions. We also propose different implementations for trading logging performance for rollback performance when this is appropriate. Finally, we revisit state-of-the-art recovery algorithms for the new context given the different latencies and synchronization control. Our results clearly indicate that current approaches for recoverability are ill-fitted for persisting data structures in the new context and it is possible to achieve low-overhead logging with customized mechanisms. Next, we focus on data evolution. We expose a simple API that allows versioning and time travel with minimal intrusiveness.We propose mechanisms for efficient and transparent cloning of Versionable data structures.This allows high concurrency since past images are returned as copies of the original data structure which remains intact. Then, we propose novel indexing techniques that significantly improve time travel performance as well as cloning with lazy schemes. We achieve a low overhead architecture by employing a mix of volatile and non-volatile data structures as well as hybrid structures that reside in both volatile and non-volatile memories. We perform an extensive evaluation of the proposed techniques and conclude that, in our context, by carefully mitigating the drawbacks of physical logging it is possible to create efficient systems for managing data evolution that are both data structure agnostic and non-intrusive.