Back to search

National Technical University of Athens (NTUA)

Accelerating virtual memory with hardware - tailored memory management

Abstract

dc:description

Since its inception in the 1960s, the paged virtual memory abstraction has become ubiquitous in computer systems. Paged virtual memory decouples the process address space from the physical memory of the system and shifts the burden of managing physical memory from userspace to the Operating System (OS). While this greatly simplifies userspace programming and enables the centralized and efficient physical memory management by the OS, it is not without trade-offs. Address Translation (AT), the process of mapping pages of virtual addresses to physical memory frames, incurs significant performance overhead. Hardware virtualization, an abstraction that has, in recent decades, become as ubiquitous as virtual memory, exacerbates this overhead. Additionally, virtualized execution creates a semantic gap between the virtual memory subsystems of the virtualized (guest) and the physical (host) OS instances, which can further affect performance. The contribution of this thesis lies in enabling the OS to take advantage of the virtual memory hardware (HW) via HW-tailored memory management to address both problems. The first part of the thesis focuses on the AT overhead of paged virtual memory and specifically on the impact of translation granularity on the performance of virtual memory, as well as on the interplay between translation granularity and OS memory management. The thesis surveys the available HW mechanisms that enable multi-grained address translation, namely large pages and TLB coalescing, and the existing OS interfaces that support these mechanisms in order to enable the OS to utilize multiple translation granules. The survey uncovers inherent limitations in these OS interfaces that hamper the effectiveness of the underlying HW mechanisms and the ability of the OS memory manager to effectively and efficiently utilize multiple translation granules. The thesis then proposes Elastic Translations (ET), a novel approach that adopts a markedly different approach regarding the way the OS interfaces with the aforementioned HW mechanisms, and enables the OS to seamlessly and optimally use and select among all HW-supported translation granules. Implementing ET in Linux for the ARMv8-A architecture and evaluating its performance on a real ARMv8-A server shows that ET unlock the full potential of the OS-assisted TLB coalescing feature of the ARMv8-A architecture and enable the OS memory manager to optimally utilize the extended range of translation granules, unlocked by coalescing, even when memory is fragmented, outperforming both state-of-practice and state-of-the-art approaches. The second part of the thesis focuses on bridging the semantic gap between the virtual memory subsystems of physical (host) and virtualized (guest) OS instances via the virtual memory hardware, i.e., the Memory Management Unit (MMU). This semantic gap becomes particularly apparent and problematic when using Virtual Machine (VM) memory snapshots to accelerate VM boot times, a technique heavily employed in latency-sensitive scenarios such as Function-as-a-Service (FaaS) serverless computing. As VM snapshots are stored as files on disk, the OS on the physical host treats the entire guest memory as a file, being oblivious to the ephemeral anonymous allocations performed by the guest memory subsystem. The thesis proposes a lightweight paravirtualized (PV) mechanism, AnonPTEs, to enlighten the host OS regarding the nature of the guest VM memory allocations by taking advantage of the virtual memory hardware. AnonPTEs piggyback on the VM AT structures, i.e., the guest VM page tables, to create a mirrored guest physical address space for the ephemeral anonymous allocations of the VM, enabling the host OS to detect such allocations and serve them using anonymous memory instead of unnecessarily fetching memory pages from the VM memory snapshot file on disk. The integration of AnonPTEs with SnapBPF, an eBPF-based state-of-the-art snapshot prefetcher for VM-sandboxed serverless functions, and its subsequent evaluation using representative serverless workloads shows that AnonPTEs eliminate gratuitous major page faults for anonymous VM allocations, thereby improving the tail latency for VM-sandboxed serverless functions.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Psomadakis, Efstratios
  • Ψωμαδάκης, Ευστράτιος

Subjects

dc:subject × 20

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60808

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Psomadakis, Efstratios; Ψωμαδάκης, Ευστράτιος. Accelerating virtual memory with hardware - tailored memory management. National Technical University of Athens (NTUA), 2025. http://hdl.handle.net/10442/hedi/60808