Full-Stack Resource Management for Predictable and Scalable Autonomous Vehicle Pipelines

Name: Dongjoo Seo

Chair: Prof. Nikil Dutt

Date: June 10, 2025

Time: 9:00 AM

Location: DBH 3011

Committee: Prof. Tony Givargis, Prof. Sang-Woo Jun

Title: Full-Stack Resource Management for Predictable and Scalable Autonomous Vehicle Pipelines

Abstract: Autonomous vehicles (AVs) rely on tightly coupled perception, planning, and control pipelines that demand both high throughput and deterministic performance from the underlying computing platform. In the first part of this dissertation, we address unpredictable memory-induced latency by introducing two complementary strategies. ProSwap leverages periodic access patterns to proactively swap pages before they incur page‐fault penalties, reducing end‐to‐end pipeline stalls by up to 26% over Linux’s default swapping. Building on this, MemScape employs reinforcement‐learning driven demotion and promotion across DRAM and high‐capacity memory tiers to avoid latency spikes—trading minimal peak performance for a 3× improvement in tail‐latency stability under dynamic AV workloads. In the second part, we tackle I/O and CPU resource management challenges that further impact AV real‐time guarantees. SPEED replaces ROS’s single‐queue architecture with a multi‐queue, context‐aware I/O scheduler, achieving an 84% reduction in rosbag latency and a 1.5× gain in throughput by aligning writebacks with application idle periods. To eliminate scheduling jitter, SCHED distills an offline‐trained reinforcement learning policy into a lightweight decision tree for in‐kernel CPU dispatch, offering up to 250× faster decision times compared to raw RL inference. Together, these full‐stack optimizations enable scalable, predictable, and safe operation of AV pipelines on commodity Linux platforms.

Bio: Dongjoo Seo received his B.E. in Computer Science and Master’s in Computer Software Engineering from Kookmin University at Seoul, Korea in 2017 and 2020, respectively. During this Master’s, Seo worked alongside Prof. Nikil Dutt at UCI as a visiting researcher where he built the simulator system for Fog computing and made a reinforcement learning-based agent for offloading the task. In 2021, Seo returned to UCI under Prof. Dutt to pursue his Ph.D. in Computer Science. His research focused on specializing OS, memory systems, and CLX. Now, Seo is a software engineer at Samsung Semiconductor, building reliable next-generation computing platforms for machine learning.