USENIX ATC2023顶会
Efficient Memory Overcommitment for I/O Passthrough Enabled VMs via Fine-grained Page Meta-data Management
Yaohui Wang, Ben Luo, Yibin Shen
摘要
In virtualization systems, guest memory overcommitment helps to improve the utilization of the host memory resource. However, the widely adopted I/O passthrough technique makes this task not intuitive since the hypervisor must avoid DMA (Direct Memory Access) failures when the I/O device accesses the guest memory. There already exist several solutions, for example, IOPF (I/O Page Fault) can fix DMA failures by allowing page fault triggered from the I/O device side, vIOMMU and coIOMMU avoid DMA failures by monitoring the DMA buffers in the guest. However, these solutions all face the performance concerns introduced by the memory backup/restore mechanism, i.e., memory swapping. Some free page based methods (e.g., Ballooning, Free Page Reporting, Hyperupcall) are free from memory swapping, but they either are not DMA-safe or introduce high guest communication overhead. In this paper, we propose V-Probe, a high-efficiency approach to achieve memory overcommitment for I/O passthrough enabled VMs. Using fine-grained page meta-data management, V-Probe allows the hypervisor to inspect and reclaim guest free pages actively and efficiently while guaranteeing DMA-safety. Experiments show that, for both memory reclamation and reallocation, the overhead of V-Probe is in the scale of microseconds, which is faster than Ballooning and IOPF based methods by two orders of magnitude. And our micro-benchmark and macro-benchmark show that V-Probe limits the performance impact of memory overcommitment to a low level.
问问这篇 Paper
智能体会读完全文。
Lune 把这篇 Paper 索引到了最后一个公式,引用它的顶会 Paper 也一样。你提问,回答直接引用原文。
引用它的顶会 Paper7
- Cloud-Native Database Systems and Unikernels: Reimagining OS Abstractions for Modern HardwareViktor Leis, Christian DietrichVLDB 2024 · 被引用 12 次
- FastIOV: Fast Startup of Passthrough Network I/O Virtualization for Secure ContainersYunzhuo Liu, Junchen Guo, Bo Jiang, Yang Song 等EuroSys 2025 · 被引用 5 次
- To PRI or Not To PRI, That's the questionYun Wang, Liang Chen, Jie Ji, Xianting Tian 等OSDI 2025 · 被引用 1 次
- VPRI: Efficient I/O Page Fault Handling via Software-Hardware Co-Design for IaaS CloudsKaijie Guo, Dingji Li, Ben Luo, Yibin Shen 等SOSP 2024 · 被引用 1 次
- Inside Out: A Paradigm Shift in VM IntrospectionDufy Teguia, Louis Duval, Teo Pisenti, Kahina Lazri 等OSDI 2026
它引用的顶会 Paper2
- SwapAdvisor: Pushing Deep Learning Beyond the GPU Memory Limit via Smart SwappingChien-Chin Huang, Gu Jin, Jinyang LiASPLOS 2020 · 被引用 161 次
- coIOMMU: A Virtual IOMMU with Cooperative DMA Buffer Tracking for Efficient Memory Management in Direct I/OKun Tian, Yu Zhang, Luwei Kang, Yan Zhao 等USENIX ATC 2020 · 被引用 18 次
相关 Paper
- HyperAlloc: Efficient VM Memory De/Inflation via Hypervisor-Shared Page-Frame AllocatorsLars Wrenger, Kenny Albes, Marco Wurps, Christian Dietrich 等EuroSys 2025 · 被引用 1 次
- Blowfish: Elastic Virtual Machine Memory for Disaggregated MemoryYulong Zhang, Yilong Luo, Diyu Zhou, Quan Chen 等OSDI 2026
- Exit-Less, Isolated, and Shared Access for Virtual MachinesKenichi Yasukata, Hajime Tazaki, Pierre-Louis AublinASPLOS 2023 · 被引用 7 次
- Compaction-Free Memory Defragmentation for Virtualization via Infinite Guest Physical Address SpacePeixin Zeng, Hao Huang, Yanqi Pan, Wen Xia 等OSDI 2026
- Fast & Safe IO Memory ProtectionBenny Rubin, Saksham Agarwal, Qizhe Cai, Rachit AgarwalSOSP 2024 · 被引用 1 次
