Digital twins and control of disk-performance in small-scale Linux-based systems

Title: Digital twins and control of disk-performance in small-scale Linux-based systems
Author: Oskar Stenberg
Date & Time: June 10th, 13:15-14:15
Location: Seminar Room M 3170-73 in the M-building, LTH
Supervisor: Björn Olofsson, Anders Blomdell
Examiner:  Bo Bernhardsson


Abstract:
Performance control in servers is important for maintaining responsiveness and efficient resource usage, particularly in smaller computing environments where resource bottlenecks are more noticeable. This creates a need for mechanisms to measure and control the performance of common bottlenecks such as disk I/O and networking.

This thesis investigates and compares two different methods for controlling disk I/O performance in virtualised Linux environments; Linux control groups (cgroups) and the QEMU virtual block layer. The goal is to prioritise high-priority workloads by throttling low-priority background tasks.

Experimental results show that both methods successfully enforce performance limits and improve the performance of prioritised workloads. In the evaluated scenarios, user-induced tasks were accelerated by up to 20% when low-priority background workloads were throttled. When measuring sequential read performance, cgroup-based limits achieved an average error of 2% compared to the set performance limits, while QEMU-based limits achieved an average error of 3%. QEMU-based limits were found to provide slightly higher accuracy and efficiency overall, while a combination of QEMU and cgroup-based limits would provide the best granularity of disk I/O control.

This thesis also investigates the deployment of a digital twin of a computer environment.The deployed twin was found to closely match the real environment, with only minor deviations in hardware specifications. To enable communication between the twin and the real environment without modifying either system, a virtual router was designed and evaluated. The router achieved a maximum throughput of approximately 9.1 Gbit/s over a 10 Gbit/s connection while introducing negligible additional latency.

In addition, several performance analysis tools were developed for monitoring CPU, memory, networking and disk performance. These include a tool for distributed data collection as well as eBPF-based debugging and tracing utilities for Linux network and storage subsystems.

Together, these components provide mechanisms for performance control, monitoring and validation of virtualised Linux environments