Data Storage Systems Commons^™

Enhancing The Performance Of Disk-Based Key-Value Stores: From Learned Index Acceleration To I/O-Efficient Hybrid Caching, Sujit Maharjan

Computer Science and Engineering Dissertations - Archive

The exponential growth of data in modern computing environments has rendered the efficient extraction of information from massive datasets a critical systemic requirement. Key-value (KV) storage systems serve as the backbone for these operations; however, their performance is consistently bottlenecked by two primary functional requirements: identifying the data's location and managing the physical cost of accessing the storage device. Data locations are typically identified via an index, while disk I/O is minimized through caching. This dissertation presents LearnedStore, TurboIndex, and ReadBooster, which break these performance bottlenecks by introducing architectural modifications to the index and cache. LearnedStore accelerates operations by adapting …

An Slo-Aware, Multi-Pronged Approach To Enhancing Resource And Energy Efficiency In Server Applications, Ning Li

Computer Science and Engineering Dissertations - Archive

Server applications operating in oversubscribed cloud environments face the dual challenges of meeting strict Quality-of-Service (QoS) requirements and improving resource and energy efficiency. As the number of user connections and workload diversity continue to grow, existing scheduling mechanisms struggle to balance QoS guarantees, fairness, resource efficiency, and power consumption. This dissertation presents a unified, cross-layer framework to address these challenges through three key contributions: AppleS, UTSLO, and REEF.

First, we propose AppleS, a user-space QoS-aware fine-grained I/O scheduling framework that delivers fair and efficient service to concurrent client connections. AppleS introduces a QoS-centric metric that guides admission control and scheduling …

Exploring Emerging Memory Technologies For Enhancing Data Center Applications, Zhen Lin

Computer Science and Engineering Dissertations - Archive

The rapid evolution of memory and storage technologies is fundamentally reshaping the design of operating systems and data management. Emerging devices such as persistent memory, NVMe SSDs, and Compute Express Link (CXL)--enabled hybrid memory modules introduce new opportunities for high-performance, cost-efficient data management, yet they also expose limitations in traditional software abstractions. File systems, originally designed for slow block-based devices, incur excessive overhead on ultra-low-latency media, while block-level caches suffer from metadata and eviction inefficiencies. Moreover, hardware-managed tiering provides transparency but restricts adaptability across workloads. These challenges highlight the need to rethink caching and tiered memory management across multiple system …

Optimizing Architecture And Software For Next-Generation Memory Systems, Lingfeng Xiang

Computer Science and Engineering Dissertations - Archive

The rapid advancement of memory technologies presents new challenges and opportunities for system software and architectural design. This dissertation investigates how to optimize modern computing systems for next-generation memory, mostly focusing on persistent memory and Compute Express Link (CXL)-based memory. First, we conduct a detailed characterization of Intel Optane DC Persistent Memory, identifying the distinct behaviors of its on-DIMM read and write buffers and analyzing their impact on application performance. These insights motivate optimizations that decouple read and write paths, revealing that random read latency—especially in pointer-chasing workloads—is a dominant performance bottleneck. Second, we present NOMAD, a page management framework …

Breaking Granularity Barriers: Overcoming I/O Abstraction Limitations For High-Performance Storage Systems, Chen Zhong

Computer Science and Engineering Dissertations - Archive

Modern storage hierarchies exhibit performance differences spanning eight orders of magnitude. Each tier in the hierarchy presents fundamentally different optimal access sizes and patterns. Most systems read and write data in fixed-size units across different tiers, typically power-of-2 sizes. It simplifies data management but also suffers from large write amplification when handling small accesses in many real-world workloads. This fundamental disconnect between fixed-size I/O design assumption and real-world usage patterns represents the core challenge.

The storage community has recognized the limitations of fixed-size I/O abstractions and is exploring alternatives, like the new emerging NVMe-KV interface and object-based storage, which aim …