Beyond the Bermuda Triangle of Contention: IOMMU Interference in Mixed Criticality Systems

TL;DR

This paper investigates how IOMMU contention causes unpredictable delays in mixed criticality systems, especially affecting small memory transactions, and highlights the importance of understanding IOMMU interference for system security and performance.

Contribution

It provides the first detailed analysis of IOMMU-induced interference effects on performance predictability in heterogeneous systems, using experimental data from Xilinx and ARM platforms.

Findings

IOMMU contention causes delays up to 1.79x for small DMA transfers.

Shared TLB structures contribute to timing unpredictability across architectures.

Interference primarily impacts small memory transactions due to translation overheads.

Abstract

As Mixed Criticality Systems (MCSs) evolve, they increasingly integrate heterogeneous computing platforms, combining general-purpose processors with specialized accelerators such as AI engines, GPUs, and high-speed networking interfaces. This heterogeneity introduces challenges, as these accelerators and DMA-capable devices act as independent bus masters, directly accessing memory. Consequently, ensuring both security and timing predictability in such environments becomes critical. To address these concerns, the Input-Output Memory Management Unit (IOMMU) plays a key role in mediating and regulating memory access, preventing unauthorized transactions while enforcing isolation and access control policies. While prior work has explored IOMMU-related side-channel vulnerabilities from a security standpoint, its role in performance interference remains largely unexplored. Moreover, many of the same architectural properties that enable side-channel leakage, such as shared TLBs, caching effects, and translation overheads, can also introduce timing unpredictability. In this work, we analyze the contention effects within IOMMU structures using the Xilinx UltraScale+ ZCU104 platform, demonstrating how their shared nature introduce unpredictable delays. Our findings reveal that IOMMU-induced interference primarily affects small memory transactions, where translation overheads significantly impact execution time. Additionally, we hypothesize that contention effects arising from IOTLBs exhibit similar behavior across architectures due to shared caching principles, such as prefetching and hierarchical TLB structures. Notably, our experiments show that IOMMU interference can delay DMA transactions by up to 1.79x for lower-size transfers on the Arm SMMUv2 implementation.