Year: 2025
Type: conference-paper
Abstract: Advances in memory-intensive areas such as machine learning and graph processing drive demand for suited memory technologies.Part of the answer lies in multi-memory architectures with fast and high-capacity memory tiers.Heterogeneous systems can be very diverse, ranging from NUMA to persistent memory and CXL-attached memory.Efficient usage requires support from the operating system (OS) with knowledge of memory performance characteristics and detailed insight into the page access behaviour of processes.Gathering statistics on the read/write-ratio of accesses and (huge) page utilisation is expensive with common methods.In this paper, we propose vmmload, a minimal virtualisation layer to monitor user space programs using a virtualisation-based page modification log (PML) for efficiently gathering runtime statistics.While PML is a hardware extension that was originally designed to track virtual machine memory access, vmmload enables observation of individual processes.vmmload monitored user space processes run with PML-enabled virtualisation, but are still able to interface with the host OS.We explore the viability of using PML without a conventional virtual machine and providing the host OS with the ability to use these extensions directly.
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Funder: Deutsche Forschungsgemeinschaft
OpenAlex ID: W4409714054