Isolation-Aware Scheduling Framework for DNN-Based End-to-End Autonomous Driving System on Tile-Based Accelerators
IEEE Transactions on Computers · 2026-09 · 75 (9) · 3226-3241
DOI: 10.1109/tc.2026.3702580 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Computers, 2026-09. Verfasst von Chenguang Zhang und 5 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben.
Abstract (Originalsprache)
With the rapid advancement of deep neural networks (DNNs) and computational capabilities, autonomous driving systems (ADS) are shifting toward DNN-based end-to-end pipelines and tile-based ADS platforms that colocate numerous small DNNs on a single chip. However, existing scheduling frameworks struggle to balance colocation efficiency, adaptive allocation, and interference isolation—a critical challenge given ADS’s strict latency constraints ( ${\leq}\,100$ ≤ 100 ms) and runtime variability. Prior solutions either adopt rigid isolation policies (reservation-based schedulers) that underutilize resources or employ fully shared policies (throughput-first schedulers) that induce excessive rescheduling overhead, both failing to ensure reliable end-to-end deadlines. We propose ADS-Tile, an isolation-aware scheduling framework that reconciles these conflicting demands. Our approach combines (1) a holistic isolation-sharing mechanism that unifies the space of interference containment and adaptive colocation, and (2) an adaptive scheduler that optimizes tile allocation using adaptive reservation techniques. Evaluations demonstrate that ADS-Tile reduces tile demand by up to 32% compared to static reservation while cutting scheduling-induced overhead from 17–44% to under 1.2% of the end-to-end latency budget, significantly advancing the feasibility of DNN-based ADS on next-generation accelerators.
Zitieren
APA
Zhang, C., Zhang, Y., Xue, C., Yin, Y., Zhang, C., & Sun, G. (2026). Isolation-Aware Scheduling Framework for DNN-Based End-to-End Autonomous Driving System on Tile-Based Accelerators. IEEE Transactions on Computers, 75(9), 3226-3241. https://doi.org/10.1109/tc.2026.3702580
Vancouver
Zhang C, Zhang Y, Xue C, Yin Y, Zhang C, Sun G. Isolation-Aware Scheduling Framework for DNN-Based End-to-End Autonomous Driving System on Tile-Based Accelerators. IEEE Transactions on Computers. 2026;75(9):3226-3241. doi:10.1109/tc.2026.3702580
BibTeX
@article{Zhang2026,
author = {Zhang, Chenguang and Zhang, Yuanpeng and Xue, Chenhao and Yin, Yihan and Zhang, Chen and Sun, Guangyu},
title = {Isolation-Aware Scheduling Framework for DNN-Based End-to-End Autonomous Driving System on Tile-Based Accelerators},
journal = {IEEE Transactions on Computers},
year = {2026},
volume = {75},
number = {9},
pages = {3226--3241},
doi = {10.1109/tc.2026.3702580}
}
EUVIMED