Abstract
dc:description.abstractThe rapid growth of data has made modern data systems increasingly storage-intensive, demanding high-performance storage solutions. In response, the shift from traditional Hard Disk Drives (HDDs) to Solid-State Drives (SSDs) has fundamentally transformed data storage, offering significantly higher performance through low-latency access, high internal parallelism, and reduced mechanical overhead. However, despite the widespread adoption of SSDs, many data-intensive systems use outdated assumptions tailored for HDDs. As a result, they fail to adapt to and harness SSD-specific characteristics such as read/write asymmetry (writes are slower than reads) and internal concurrency (SSDs can read or write multiple pages in parallel without hurting latency). Without exploiting internal concurrency, the device can remain vastly underutilized. To fully unlock SSD potential and ensure both optimal performance and device longevity, systems need to be redesigned with SSD-aware strategies.This thesis bridges the gap between SSD properties and data system optimizations by more faithful modeling that captures key properties of SSDs: asymmetry and concurrency. First, we propose a Parametric I/O Model (PIO) that quantifies SSD properties (read/write asymmetry and concurrency) and provides guidelines for SSD-aware algorithm design. Next, we develop ACE, an asymmetry/concurrency-aware bufferpool manager, which batches writes to write them in parallel and performs parallel prefetching to utilize the full bandwidth of the underlying SSD device. We then introduce CAVE, an out-of-core graph processing engine that optimally schedules independent graph I/O operations to maximize SSD read parallelism by exploring multiple paths concurrently. We implement parallelized variants of DFS and BFS that exploit storage parallelism to offer the same algorithm guarantees with better average performance. These implementations also serve as the building block of other graph algorithms (weekly connected components, pagerank, random walk). Finally, we present ReStore, the first data-migration policy for multi-tiered architectures that uses storage concurrency and asymmetry when modeling the state of the devices. ReStore dynamically places data across tiers based on the device properties and workload characteristics to optimize performance and cost. Through extensive benchmarking and experiments on real SSDs, we showcase the benefit of better storage modeling in algorithm and system design to improve performance. Our evaluation shows that ACE bufferpool manager improves runtime performance by up to 32% for transactional workloads, CAVE graph manager is up to 10x faster than state-of-the-art baselines for various traversal algorithms and ReStore tiered data migration achieves up to 2.2x lower runtime and causes up to 10x fewer migrations. The findings in this thesis contribute to the broader field of storage system design, influencing databases, file systems and storage-intensive applications by faithfully leveraging SSD properties for higher performance.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Papon, Tarikul Islam
- Advisor dc:contributor.advisor
-
- Athanassoulis, Manos
- ORCID dc:identifier.orcid
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53075