Treffer: Performance optimization and research of data center-level IOMMU based on RISC-V.

Title:
Performance optimization and research of data center-level IOMMU based on RISC-V.
Authors:
Zhao, Shuntong1 (AUTHOR), Hong, Mingsen1 (AUTHOR), Wang, Xiaohui1 (AUTHOR) wangxiaohui@cmss.chinamobile.com, Wang, Dongxu1 (AUTHOR), Liu, Yanan1 (AUTHOR)
Source:
Procedia Computer Science. 2025, Vol. 266, p923-935. 13p.
Database:
Supplemental Index

Weitere Informationen

From the perspectives of architecture and market, IaaS (Infrastructure as a Service) remains the core form of cloud data centers, supporting PaaS (Platform as a Service) and SaaS (Software as a Service). The core feature of IaaS is virtualization, which maximizes resource utilization and economic efficiency through the reuse of underlying hardware. With the growing complexity of data center tasks and the increasing demand for computing power and network processing capabilities, efficient and secure data interaction between I/O devices and processors has become increasingly important. As a result, higher requirements are being placed on IOMMU technology within the field of cloud computing. In response to these challenges, this paper proposes a RISC-V architecture IOMMU system design for cloud data centers. Compared to the IOMMU v1.0.0 specification architecture, this design adopts a distributed IOATC module and an innovative Walk Cache IOATS module. Through UVM simulation verification and testing, it has been shown that the IOATC module's direct address hit leads to nearly 50 times performance acceleration. Furthermore, the IOATS module supporting the distributed IOATC module significantly improves address translation efficiency without introducing any delay. The Walk Cache structure increases translation efficiency by about 30% compared to the PTW cache mode. In actual data center tests, network virtualization data testing was performed on a mobile cloud BC-Linux OS, equipped with the self-developed high-speed peripheral Panshi DPU4.0. The results indicate a performance improvement of about 2.6 times compared to the traditional Virtio-Net software implementation. [ABSTRACT FROM AUTHOR]