New architectures for non-volatile memory technologies

Over the last few decades, there has been a drastic shift toward applications that need to store and operate on a large volume of data and, at the same time, provide a high quality of service to millions of users. One popular example of these types of applications is online services. For these applications, Non-Volatile Memory (NVM) technologies, such as Intel 3D XPoint, are particularly interesting. NVM has high density, it is byte addressable, and offers an attractive balance between performance and non-volatility to store data. However, the success of NVM depends on the availability of architectures that gracefully facilitate NVM’s integration to current systems and take advantage of its properties. This dissertation develops computing architectures which ease the adoption of NVMs in contemporary systems. It addresses the challenges of utilizing NVMs both (a) in single node multicore systems, as well as (b) in multi-node distributed systems. For single node systems, this thesis first addresses the problem of data placement in systems that use both DRAM and NVM in a hybrid main memory hierarchy, and introduces PageSeer, which leverages page walks to predict forthcoming requests to main memory, and swaps data between DRAM and NVM to achieve high performance. Next, it examines NVM as a cache replacement and suggests Cloak, a technique that can hide the increased latency of NVM and take advantage of its high density when used as a last level cache by leveraging address translation to identify page reuse. Next, this thesis addresses the problem of programming for NVMs. Easy to use programming frameworks that dynamically detect persistent objects suffer from low performance. Thus, this thesis proposes P-INSPECT, which uses hardware to accelerate such programming frameworks. For distributed systems, this thesis introduces the concept of Distributed Data Persistency (DDP), which is the binding of consistency models with different memory persistency models. To support DDP models, it develops low-latency protocols that provide a variety of different options for persisting and replicating data in distributed systems with NVMs. It also analyzes the advantages and disadvantages of each of these options. Moreover, it proposes HADES, which adds hardware support for distributed transactional consistency, to bypass the costly software abstractions that increase client request latency and programming complexity, and assesses the interaction of HADES with the different persistency models. It also utilizes the network hardware to speed up data persist operations. Overall, this thesis proposes solutions to the different NVM challenges. These solutions can pave the way for the wide integration of NVM in future systems.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Kokolis, Apostolos

Contributors dc:contributor

• Torrellas, Josep
• Xu, Tianyin
• Huang, Jian
• Padua, David
• Kalamatianos, John
• Wenisch, Thomas F.
• Nagarajan, Vijay

• Non-Volatile Memory
• Hybrid Memory Systems
• Distributed Architectures
• Memory Persistency
• Distributed Transactions
• Hardware for Programmability
• Virtual Memory

Statement dc:rights

• Copyright 2022 Apostolos Kokolis

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115720

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115720