Master thesis : Load time memory deduplication for unikernels

Abstract

Memory deduplication is a crucial technique in virtualized environments to optimize memory usage and enhance resource allocation. Unikernels, lightweight single-purpose virtual machine images, present unique challenges in memory management due to their specialized design. This thesis explores the intricacies of memory deduplication in unikernels and introduces the Daemonless Kernel Same-page Merging (DKSM) mechanism, a novel approach to memory deduplication tailored for unikernel environments.

The primary objective of this thesis is to develop a load-time memory deduplication technique for Unikraft unikernels in the Linux kernel. DKSM is an alternative to Red Hat’s Kernel Same-page Merging (KSM), but it focuses exclusively on merging read-only pages. This distinction allows DKSM to operate during unikernel startup, reducing the ever-present processing overhead found in KSM.

One of DKSM’s notable contributions is its mitigation of the slow convergence issue, an inherent limitation of KSM. DKSM also demonstrates better performance on long-lived unikernels, while benefiting from the same memory gain as KSM. Additionally, due to its daemonless nature, DKSM offers the advantage of operating without an additional processor core, making it more resource-efficient compared to the traditional KSM approach.

This thesis presents an in-depth analysis of Daemonless Kernel Same-page Merging’s implementation, its integration into the Linux kernel, and its validation through performance testing using unikernels executed within the Firecracker hypervisor. Results demonstrate DKSM’s efficiency in memory deduplication and its potential to improve memory utilization in unikernel environments. The study concludes by discussing limitations and potential avenues for future research.