TY - BOOK

T1 - Automated tailoring of system software stacks

AU - Ziegler, Andreas

PY - 2023

Y1 - 2023

N2 - In many industrial sectors, device manufacturers are moving away from expensive special-purpose hardware units and consolidate their systems on commodity hardware. As part of this change, developers are enabled to run their applications on general-purpose operating systems like Linux, which already supports thousands of different devices out of the box and can be used in a wide range of target scenarios. Furthermore, the Linux ecosystem allows them to integrate existing implementations of standard functionality in the form of shared libraries. However, as the libraries and the Linux kernel are designed as generic building blocks in order to support as many applications as possible, they cannot make assumptions about specific use cases for a single-purpose device. This generality leads to unnecessary overheads in narrowly defined target scenarios, as unneeded components do not only take up space on the target system but have to be maintained over the lifetime of the device as well. While the Linux kernel provides a configuration system to disable unneeded functionality like device drivers, determining the required features from over 16000 options is an infeasible task. Even worse, most shared libraries cannot be customized even though only around 10 percent of their functions are ever used by applications. In this thesis, I present my approaches for the automated identification and removal of unnecessary components in all layers of the software stack. As the configuration system is an integral part of the Linux kernel, we embrace its presence and automatically generate custom-fitted configurations for observed target scenarios with the help of an extracted variability model. For the much more diverse realm of shared libraries, with different programming languages, build systems, and a lack of configurability, I demonstrate a different approach. By identifying individual functions as logically distinct units, we construct a symbol-level dependency graph across the applications and all their required libraries. We then remove unneeded code at the binary level and rearrange the remaining parts to take up minimal space in the binary file by formulating their placement as an optimization problem. To lower the number of unnecessary updates to unused components in a deployed system, I lastly present an automated method to determine the impact of software changes on a target scenario and provide guidance for developers on whether they need to update their systems. Applying these techniques to different target systems, I demonstrate that we can disable up to 87 percent of configuration options in a Debian Linux kernel, shrink the size of an embedded OpenWrt kernel by 59 percent, and speed up the boot process of the embedded system by 21 percent. As part of the shared library tailoring process, we can remove 13060 functions from all libraries in OpenWrt and reduce their total size by 31 percent. In the memcached Docker container, we identify 381 entirely unneeded shared libraries and shrink the container image size by 82 percent. An analysis of the development history of two large library projects over the course of more than two years further shows that between 68 and 82 percent of all changes are not required for an OpenWrt appliance, reducing the number of patch days by up to 69 percent. These results demonstrate the broad applicability of our automated methods for both the Linux kernel and shared libraries to a wide range of scenarios. From embedded systems to server applications, custom-tailored system software stacks contribute to the reduction of overheads in space and time.

AB - In many industrial sectors, device manufacturers are moving away from expensive special-purpose hardware units and consolidate their systems on commodity hardware. As part of this change, developers are enabled to run their applications on general-purpose operating systems like Linux, which already supports thousands of different devices out of the box and can be used in a wide range of target scenarios. Furthermore, the Linux ecosystem allows them to integrate existing implementations of standard functionality in the form of shared libraries. However, as the libraries and the Linux kernel are designed as generic building blocks in order to support as many applications as possible, they cannot make assumptions about specific use cases for a single-purpose device. This generality leads to unnecessary overheads in narrowly defined target scenarios, as unneeded components do not only take up space on the target system but have to be maintained over the lifetime of the device as well. While the Linux kernel provides a configuration system to disable unneeded functionality like device drivers, determining the required features from over 16000 options is an infeasible task. Even worse, most shared libraries cannot be customized even though only around 10 percent of their functions are ever used by applications. In this thesis, I present my approaches for the automated identification and removal of unnecessary components in all layers of the software stack. As the configuration system is an integral part of the Linux kernel, we embrace its presence and automatically generate custom-fitted configurations for observed target scenarios with the help of an extracted variability model. For the much more diverse realm of shared libraries, with different programming languages, build systems, and a lack of configurability, I demonstrate a different approach. By identifying individual functions as logically distinct units, we construct a symbol-level dependency graph across the applications and all their required libraries. We then remove unneeded code at the binary level and rearrange the remaining parts to take up minimal space in the binary file by formulating their placement as an optimization problem. To lower the number of unnecessary updates to unused components in a deployed system, I lastly present an automated method to determine the impact of software changes on a target scenario and provide guidance for developers on whether they need to update their systems. Applying these techniques to different target systems, I demonstrate that we can disable up to 87 percent of configuration options in a Debian Linux kernel, shrink the size of an embedded OpenWrt kernel by 59 percent, and speed up the boot process of the embedded system by 21 percent. As part of the shared library tailoring process, we can remove 13060 functions from all libraries in OpenWrt and reduce their total size by 31 percent. In the memcached Docker container, we identify 381 entirely unneeded shared libraries and shrink the container image size by 82 percent. An analysis of the development history of two large library projects over the course of more than two years further shows that between 68 and 82 percent of all changes are not required for an OpenWrt appliance, reducing the number of patch days by up to 69 percent. These results demonstrate the broad applicability of our automated methods for both the Linux kernel and shared libraries to a wide range of scenarios. From embedded systems to server applications, custom-tailored system software stacks contribute to the reduction of overheads in space and time.

U2 - 10.15488/15610

DO - 10.15488/15610

M3 - Dissertation

CY - Hannover

ER -