Available at: https://digitalcommons.calpoly.edu/theses/3037
Date of Award
6-2025
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
John Bellardo
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Modern web servers face unprecedented demands for high throughput and low latency [10] [11]. Yet, even the state-of-the-art optimizations often fail under heavy workloads on existing infrastructure. Despite advancements in hardware, communication between applications and the kernel remains a critical bottleneck.
Industry surveys reveal that over 50% of production servers still rely on traditional epoll-based architectures [5]. This research aims to investigate and characterize Linux’s new io_uring subsystem, which has the potential to overcome these challenges. Through controlled load testing of various existing architectures like event-driven, multi-process, and multi-threaded architectures (including other commercial servers), we demonstrate how io_uring can help achieve better benchmarks both with and without cached data. Our methodology employs controlled load testing while maintaining constant network bandwidth as the primary bottleneck, with detailed measurement of key system metrics including syscalls, memory usage, disk I/O performance, and CPU utilization patterns.
This work provides empirical guidance for migrating performance-critical services to io_uring, while also outlining remaining challenges for future work.