Morsels: Explicit Virtual Memory Objects

DIMES Workshop
Morsels: Explicit Virtual Memory Objects

Alexander Halbuer, Christian Dietrich, Florian Rommel, Daniel LohmannProceedings of the 1st Workshop on Disruptive Memory SystemsAssociation for Computing Machinery2023.
PDF Details Slides 10.1145/3609308.3625267 [BibTex]

×

BibTex Entry

@inproceedings{halbuer:23:dimes,
author = {Alexander Halbuer and Christian Dietrich and Florian Rommel and Daniel Lohmann},
booktitle = {Proceedings of the 1st Workshop on Disruptive Memory Systems},
date = {2023-10},
doi = {10.1145/3609308.3625267},
entrysubtype = {Workshop},
location = {Koblenz, Germany},
month = {October},
publisher = {Association for Computing Machinery},
series = {DIMES '23},
title = {Morsels: Explicit Virtual Memory Objects},
url = {https://doi.org/10.1145/3609308.3625267},
year = {2023},
}

Close

Abstract

The tremendous growth of RAM capacity – now exceeding multiple terabytes – necessitates a reevaluation of traditional memory-management methods, which were developed when resources were scarce. Current virtual-memory subsystems handle address-space regions as sets of individual 4-KiB pages with demand paging and copy-on-write, resulting in significant management overhead. Although huge pages reduce the number of managed entities, they induce internal fragmentation and have a coarse copy granularity.

To address these problems, we introduce Morsels, a novel virtual-memory–management paradigm that is purely based on hardware data structures and enables the efficient sharing of virtual-memory objects between processes and devices while being well suited for non-volatile memory. Our benchmarks show that Morsels reduce the mapping time for a 6.82-GiB machine-learning model by up to 99.8 percent compared to conventional memory mapping in Linux.