Files in this item

FilesDescriptionFormat

application/pdf

application/pdfSHULL-DISSERTATION-2020.pdf (1MB)
(no description provided)PDF

Description

Title:Making non-volatile memory programmable
Author(s):Shull, Thomas Edward
Director of Research:Torrellas, Josep
Doctoral Committee Chair(s):Torrellas, Josep
Doctoral Committee Member(s):Huang, Jian; Padua, David; Larus, James; Swanson, Steven
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):non-volatile memory
java
instruction set architecture
memory models
Abstract:Byte-addressable, non-volatile memory (NVM) is emerging as a revolutionary memory technology that provides persistence, near-DRAM performance, and scalable capacity. By using NVM, applications can directly create and manipulate durable data in place without the need for serialization out to SSDs. Ideally, through NVM, persistent applications will be able to maintain crash-consistency at a minimal cost. However, before this is possible, improvements must be made at both the hardware and software level to support persistent applications. Currently, software support for NVM places too high of a burden on the developer, introducing many opportunities for mistakes while also being too rigid for compiler optimizations. Likewise, at the hardware level, too little information is passed to the processor about the instruction-level ordering requirements of persistent applications; this forces the hardware to require the use of coarse fences, which significantly slow down execution. To help realize the promise of NVM, this thesis proposes both new software and hardware support that make NVM programmable. From the software side, this thesis proposes a new NVM programming model which relieves the programmer from performing much of the accounting work in persistent applications, instead relying on the runtime to perform error-prone tasks. Specifically, within the proposed model, the user only needs to provide minimal markings to identify the persistent data set and to ensure data is updated in a crash-consistent manner. Given this new NVM programming model, this thesis next presents an implementation of the model in Java. I call my implementation AutoPersist and build my support into the Maxine research Java Virtual Machine (JVM). In this thesis I describe how the JVM can be changed to support the proposed NVM programming model, including adding new Java libraries, adding new JVM runtime features, and augmenting the behavior of existing Java bytecodes. In addition to being easy-to-use, another advantage of the proposed model is that it is amenable to compiler optimizations. In this thesis I highlight two profile-guided optimizations: eagerly allocating objects directly into NVM and speculatively pruning control flow to only include expected-to-be taken paths. I also describe how to apply these optimizations to AutoPersist and show they have a substantial performance impact. While designing AutoPersist, I often observed that dependency information known by the compiler cannot be passed down to the underlying hardware; instead, the compiler must insert coarse-grain fences to enforce needed dependencies. This is because current instruction set architectures (ISA) cannot describe arbitrary instruction-level execution ordering constraints. To fix this limitation, I introduce the Execution Dependency Extension (EDE), and describe how EDE can be added to an existing ISA as well as be implemented in current processor pipelines. Overall, emerging NVM technologies can deliver programmer-friendly high performance. However, for this to happen, both software and hardware improvements are necessary. This thesis takes steps to address current the software and hardware gaps: I propose new software support to assist in the development of persistent applications and also introduce new instructions which allow for arbitrary instruction-level dependencies to be conveyed and enforced by the underlying hardware. With these improvements, hopefully the dream of programmable high-performance NVM is one step closer to being realized.
Issue Date:2020-07-15
Type:Thesis
URI:http://hdl.handle.net/2142/108493
Rights Information:2020 Thomas Shull
Date Available in IDEALS:2020-10-07
Date Deposited:2020-08


This item appears in the following Collection(s)

Item Statistics