Thesis title: Accelerating Virtual Memory with Hardware-Tailored Memory Management

Abstract: The dissertation addresses inefficiencies in modern virtual memory systems by designing hardware-tailored mechanisms and policies for the Operating System (OS) memory manager, that improve address translation performance and reduce virtualization-induced overheads. It first introduces Elastic Translations (ET), a holistic extension to the OS memory manager that exploits OS-assisted TLB coalescing, available on modern architectures, such as ARMv8-A tand RISC-V, to transparently and efficiently use multiple translation granules. Through coalescing-aware demand paging and asynchronous promotions, and hardware-assisted TLB-miss sampling, ET significantly improves performance for both native and virtualized workloads, even under fragmentation. The second part of the thesis focuses on serverless computing and proposes AnonPTEs, a lightweight paravirtualized mechanism that bridges the semantic gap between the virtual memory subsystems of guest and host OS, when executing microVM-sandboxed serverless functions from VM memory snapshot files. By conveying the nature of the guest OS memory allocations to the host, AnonPTEs eliminates unnecessary I/O and major page faults for the ephymeral anonymous allocations of the guest OS, substantially improving tail latency and overall system performance. Collectively, the thesis demonstrates that aligning OS memory management with underlying hardware capabilities unlocks substantial performance gains.

Supervisor: Professor Georgios Goumas

PhD Student: Stratos Psomadakis