Wei Zhang

Wei Zhang

Auto-generated profile, based on publicly available information

Wei Zhang

Brown University, Providence, USA

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
29 June 2025
Verified
Conference Paper
Open access
SJShixin Ji
Shixin Ji
XCXingzhen Chen
Xingzhen Chen
JZJinming Zhuang
Jinming Zhuang
WZWei Zhang
Wei Zhang
ZYZhuoping Yang
Zhuoping Yang
PZPeipei Zhou
Peipei Zhou

27 February 2025
Verified
Conference Paper
Open access
SJShixin Ji
Shixin Ji
XCXingzhen Chen
Xingzhen Chen
WZWei Zhang
Wei Zhang
ZYZhuoping Yang
Zhuoping Yang
JZJinming Zhuang
Jinming Zhuang
PZPeipei Zhou
Peipei Zhou
Real-time requirements are essential for safety-critical systems such as autonomous driving and robotics. In these systems, schedulability - ensuring every task completes before its deadline - is critical, as missing even a single deadline can lead to catastrophic consequences. Unlike Quality-of-Service (QoS) or service-level agreement (SLA) metrics, where occasional deadline misses are tolerable, time-predictable safety-critical systems demand strict adherence to deadlines. Existing multi-task scheduling frameworks, optimized for QoS/SLA, fail to meet these requirements, even when incorporating real-time techniques like preemption and scheduling. In FPGA-based systems, accelerators are widely used to handle real-time computational tasks. While existing accelerators focus on low latency and high throughput, these optimizations are insufficient when multiple tasks share a single accelerator, as low latency alone does not ensure schedulability. For example in Fig.1(a), a system with tasks of 20 ms and 400 ms periods, requiring 10 ms and 200 ms execution times, will miss deadlines on a 1 TOPS accelerator. Even increasing throughput to 10 TOPS (Fig.1(b)) without proper scheduling fails to resolve the issue. Furthermore, without a theoretical analysis, designers cannot determine whether the system after latency optimization can meet the deadline. As a result, these latency-optimized accelerators can not be safely used in time-predictable, safety-critical real-time systems. To address these challenges, we propose the ART framework: Accelerator Customization for Real-Time Systems. ART enables FPGA accelerators to support Earliest Deadline First (EDF) scheduling and limited preemption mechanisms, specifically targeting real-time safety-critical systems. Alongside hardware support, ART provides a software framework to analyze schedulability and balance between schedulability and performance. As shown in Figure 1(c), a 1 TOPS accelerator with ART can satisfy the schedulability requirement. Whereas the latency-optimized accelerator without ART in Figure 1(b) can not satisfy and thus is not viable in this scenario. Our contributions are summarized as follows: Necessity analysis of scheduling and preemption mechanism: We conduct a detailed case study to understand the necessity of scheduling and preemption. ART hardware framework and accelerator system: We propose the ART hardware framework and accelerator system for EDF scheduling and limited preemption on FPGA. ART can transform a baseline HLS accelerator to support these features. ART software framework: To guarantee the schedulability in real-time scenarios, we propose the ART software framework. We migrate the schedulability analysis and optimal preemption point placements to hardware acceleration, explain their effectiveness and package these algorithms into an automated framework. ART implementations and experiments: We implement and evaluate the ART accelerator system in AMD Zynq ZCU102, and Versal VCK190 platforms. The onboard experiment shows that ART guarantees the schedulability whereas the corresponding baseline accelerators can not. ART is lightweight, utilizing < 0.4% of resources on ZCU102 and < 0.1% on VCK190. Necessity analysis of scheduling and preemption mechanism: We conduct a detailed case study to understand the necessity of scheduling and preemption. ART hardware framework and accelerator system: We propose the ART hardware framework and accelerator system for EDF scheduling and limited preemption on FPGA. ART can transform a baseline HLS accelerator to support these features. ART software framework: To guarantee the schedulability in real-time scenarios, we propose the ART software framework. We migrate the schedulability analysis and optimal preemption point placements to hardware acceleration, explain their effectiveness and package these algorithms into an automated framework. ART implementations and experiments: We implement and evaluate the ART accelerator system in AMD Zynq ZCU102, and Versal VCK190 platforms. The onboard experiment shows that ART guarantees the schedulability whereas the corresponding baseline accelerators can not. ART is lightweight, utilizing < 0.4% of resources on ZCU102 and < 0.1% on VCK190.

22 June 2026
Verified
Conference Paper
Open access
SJShixin Ji
Shixin Ji
JZJinming Zhuang
Jinming Zhuang
ZYZhuoping Yang
Zhuoping Yang
XCXingzhen Chen
Xingzhen Chen
WZWei Zhang
Wei Zhang
PZPeipei Zhou
Peipei Zhou