Search for: All records

This paper presents a real-time locking protocol whose design was motivated by the goal of enabling safe GPU sharing in time-sliced component-based systems. This locking protocol enables a GPU to be shared concurrently across, and utilized within, isolated components with predictable execution times. It relies on a novel resizing technique where GPU work is dimensioned on-the-fly to run on partitions of an NVIDIA GPU. This technique can be applied to any component that internally utilizes global CPU scheduling. The proposed locking protocol enables increased GPU parallelism and reduces GPU capacity loss with analytically provable benefits. 

The goal of a real-time locking protocol is to reduce any priority-inversion blocking (pi-blocking) a task may incur while waiting to access a shared resource. For mutual-exclusion sharing on an m-processor platform, the best existing lower bound on per-task pi-blocking under suspension-oblivious analysis is a (trivial) lower bound of (m-1) request lengths under any job-level fixed-priority (JLFP) scheduler. Surprisingly, most asymptotically optimal locking protocols achieve a per-task pi-blocking upper bound of (2m-1) request lengths under JLFP scheduling, even though a range of very different mechanisms are used in these protocols. This paper closes the gap between these existing lower and upper bounds by establishing a lower bound of (2m-2) request lengths under global fixed-priority (G-FP) and global earliest-deadline-first (G-EDF) scheduling. This paper also shows that worst-case per-task pi-blocking can be arbitrarily close to (2m-1) request lengths for locking protocols that satisfy a certain property that is met by most (if not all) existing locking protocols. These results imply that most known asymptotically optimal locking protocols are almost truly optimal (not just asymptotic) under G-FP and G-EDF scheduling. 

Machine-learning (ML) technology has been a key enabler in the push towards realizing ever more sophisticated autonomous-driving features. In deploying such technology, the automotive industry has relied heavily on using "black-box" software and hardware components that were originally intended for non-safety-critical contexts, without a full understanding of their real-time capabilities. A prime example of such a component is CUDA, which is fundamental to the acceleration of ML algorithms using NVIDIA GPUs. In this paper, evidence is presented demonstrating that CUDA can cause unbounded task delays. Such delays are the result of CUDA’s usage of synchronization mechanisms in the POSIX thread (pthread) library, so the latter is implicated as a delay-prone component as well. Such synchronization delays are shown to be the source of a system failure that occurred in an actual autonomous vehicle system during testing at WeRide. Motivated by these findings, a broader experimental study is presented that demonstrates several real-time deficiencies in CUDA, the glibc pthread library, Linux, and the POSIX interface of the safety-certified QNX Operating System for Safety. Partial mitigations for these deficiencies are presented and further actions are proposed for real-time researchers and developers to integrate more complete mitigations.