Practical DB-OS Co-Design for Modern High-Performance Database Engines
Name
zhou-xinjing-phd-eecs-2026-thesis.pdf
Size
6.01 MB
Format
Adobe PDF
Checksum (MD5)
38db14b5794cc7c744696a069d422212
Author(s)
Zhou, Xinjing
Advisor(s)
Stonebraker, Michael
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Database management systems (DBMSs) rely on operating systems (OSes) to access hardware resources, but the abstractions exposed by general-purpose OSes are often poorly matched to database workloads. DBMSs are optimized to process data-intensive workloads, while OSes provide interfaces that emphasize security, protection, and fair multiplexing of hardware resources. This long-standing mismatch has become more severe as modern hardware evolves rapidly. Networks and storage devices have improved by orders of magnitude, while single-core CPU performance has improved much more slowly. As a result, operating systems increasingly dominate query and transaction execution time. Existing approaches such as full kernel bypass, custom kernels, and kernel modules can reduce this overhead, but they often sacrifice key practicality properties, including compatibility, security, deployability, and maintainability.
This thesis argues for practical DB-OS co-design, replacing poorly matched OS mechanisms on the database critical path while preserving these practicality properties. It begins with a whole-stack measurement study of memory-optimized OLTP systems, which identifies the OS/kernel stack as a major source of cost under modern hardware and workloads. Guided by the principle of specializing only the mechanisms that limit performance, and doing so at the lowest sufficient privilege level, the thesis develops three systems. Tux is a database-aware networking stack that achieves kernel-bypass-class performance without requiring user-space NIC drivers or exclusive device ownership. Calico is a DBMS-controlled buffer pool that matches hardware-accelerated in-memory translation while preserving DBMS control over eviction and scalable I/O. Libdbos leverages hardware virtualization to safely expose privileged mechanisms to the database process, enabling instantaneous virtual-memory snapshots and efficient buffer management.
Together, these systems show that the DB-OS interface can be redesigned around database semantics across communication, memory management, and privilege boundaries. The result is a practical path to large end-to-end performance gains without giving up the practicality properties required by production database systems.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link