Real-Time Scheduling of TrustZone-enabled DNN Workloads
Real-Time Scheduling of TrustZone-enabled DNN Workloads
复制标题
DOI：
10.1145/3560826.3563386
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Mohammad Fakhruddin Babar;M. Hasan
中科院分区：
文献类型：
--
作者：
Mohammad Fakhruddin Babar;M. Hasan
嵌入式设备中有限的资源通常会阻碍计算繁重的机器学习过程的执行。在运行深度神经网络（DNN）工作负载的同时保持模型参数的完整性，并且不影响实时应用的时间约束，是一个具有挑战性的问题。尽管像ARM TrustZone这样的安全区域可以确保应用程序的完整性，但由于额外的资源和时间限制，现成的实现对于DNN工作负载通常是不可用的——尤其是那些有实时要求的工作负载。本文提出了一个实时调度框架，可以在启用trustzone的安全飞地内执行资源密集型DNN工作负载。我们的方法通过融合多个任务的多个层并在保留实时授权的同时在enclave内一起运行它们来减少资源开销。我们推导出数学条件，使设计人员能够测试在启用trustzone的系统中部署DNN工作负载的可行性。我们与标准固定优先级实时调度器的比较表明，在更高的利用率（例如，bbb80 %）场景中，我们可以调度多达21.33%的任务集。
Limited resources in embedded devices often hinder the execution of computation-heavy machine learning processes. Running deep neural network (DNN) workloads while preserving the integrity of the model parameters and without compromising temporal constraints of real-time applications, is a challenging problem. Although secure enclaves such as ARM TrustZone can ensure the integrity of applications, off-the-shelf implementations are often infeasible for DNN workloads - especially those with real-time requirements - due to additional resource and temporal constraints. This paper presents a real-time scheduling framework that enables the execution of resource-intensive DNN workloads inside TrustZone-enabled secure enclaves. Our approach reduces the resource overhead by fusing multiple layers of multiple tasks and running them all together inside the enclaves while retaining real-time grantees. We derive mathematical conditions that will allow the designer to test the feasibility of deploying DNN workload in a TrustZone-enabled system. Our comparisons with a standard fixed-priority real-time scheduler show that we can schedule up to 21.33% more tasksets in higher utilization (e.g., > 80%) scenarios.