Scholay

学术搜索 · AI 审稿 · LaTeX 协作

Sweet or Sour CHERI: Performance Characterization of the Arm Morello Platform

作者:Xiaoyang Sun, Jeremy Singer, Zheng Wang · 年份:2025 · DOI:10.1109/iiswc66894.2025.00042 · 被引用次数:2 · 研究领域:Security and Verification in Computing、Radiation Effects in Electronics、Cryptographic Implementations and Security

Capability Hardware Enhanced RISC Instructions (CHERI) is an emerging hardware approach to memory safety that enforces strong spatial and temporal protections. This paper presents the most comprehensive performance evaluation of CHERI to date. Using on-chip performance monitoring counters (PMCs) on the CHERI-enabled Arm Morello platform, we analyze 20 C/C++ applications, including SPEC CPU2017, a SQL database engine, a JavaScript engine, and a large language model inference framework, across three CHERI Application Binary Interfaces (ABIs). We find that CHERI overheads range from negligible to 1.65×, with the highest costs in pointer-intensive and memory-sensitive workloads, largely due to increased memory traffic and higher L1/L2 cache pressure from 128-bit capabilities. Importantly, our projections suggest that modest microarchitectural improvements could significantly reduce these costs, enabling CHERI to deliver memory safety with minimal performance impact. We hope these findings offer timely evidence to guide the development of future architectures that combine strong memory security with high performance.