MTTZ: A Multithreading Framework for Parallel Processing in ARM TrustZone

Abstract

As computing systems evolve toward large-scale multicore processors, achieving efficient parallel processing in trusted execution environments (TEEs), such as ARM TrustZone, has become a significant challenge. The current single-threaded model of TrustZone limits multicore parallel computing capabilities, and existing multithreading solutions suffer from high thread creation overhead. Furthermore, the limited standard library support and absence of dynamic library loading capabilities in TEE significantly hinder the development of complex secure applications, forcing developers to implement low-level primitives and restricting code reusability. This article proposes MTTZ, a novel multithreading framework for ARM TrustZone that addresses these limitations through three key components: 1) a secure and efficient thread monitor at the TEE driver layer, which interacts with the TEE to manage multithread creation and significantly reduces creation overhead; 2) a comprehensive parallel and concurrency mechanism, achieving low-overhead multitask processing through thread-level parallelism and lightweight coroutine concurrency; and 3) a cross-world library loading method, enabling secure loading of dynamic libraries with dramatic size reductions. We implement a prototype of MTTZ on off-the-shelf ARM platforms and perform a comprehensive set of experiments. The experimental results show that MTTZ reduces thread creation overhead by 42% compared to the state-of-the-art work and achieves up to 98.97% binary size reduction through dynamic library loading. Comprehensive evaluations on real devices demonstrate that MTTZ achieves near-linear scalability with over 90% parallel efficiency, enabling up to $3.5\times $ speedup on TensorFlow Lite and $2.92\times $ speedup on small language model (SLM) inference while maintaining approximately 10% average overhead compared to native Linux.