A dynamic task partitioning method for CPU–FPGA heterogeneous systems for computing intensive workloads
Journal of Vibration and Control · 20.11.2026
DOI: 10.1177/10775463261508672 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Vibration and Control, online erschienen am 20.11.2026. Verfasst von Qinghuai Liu und 1 weiteren Autor:innen. Die Arbeit nennt 17 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Fixed CPU–FPGA workload ratios can become inefficient when task characteristics or available resources change. A window-based partitioning framework is examined for divisible continuous workloads with resident FPGA kernels. The framework combines an execution-time model, a discrete search over the FPGA workload ratio, exponentially weighted moving average updates, and hysteresis with a minimum-gain test. The model consistently treats CPU execution and the complete FPGA branch as concurrent, while transfers and FPGA computation are serialized within the FPGA branch. An affine special case admits an analytical balance solution; discrete search is retained for non-affine costs and feasibility constraints. In the reported ZCU102 timing summaries, the method has an equally weighted mean latency of 16.47 ms across three workload scales and a pooled speedup of 1.94× over CPU-only execution. The 1.0% latency difference from the offline-calibrated static baseline is descriptive and has not been shown to be statistically significant. During the reported disturbance stage, prediction MAPE is 9.6%, compared with 18.7% without feedback; reported ratio changes decrease from 86 to 31 with stability control. These summaries indicate improved prediction and fewer ratio changes, but do not establish a disturbance-stage latency benefit, superiority over existing dynamic runtimes, or negligible scheduling overhead. Such claims require additional measurements and reproducibility records.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Liu, Q., & Tao, W. (2026). A dynamic task partitioning method for CPU–FPGA heterogeneous systems for computing intensive workloads. Journal of Vibration and Control. https://doi.org/10.1177/10775463261508672
Vancouver
Liu Q, Tao W. A dynamic task partitioning method for CPU–FPGA heterogeneous systems for computing intensive workloads. Journal of Vibration and Control. 2026. doi:10.1177/10775463261508672
BibTeX
@article{Liu2026,
author = {Liu, Qinghuai and Tao, Weiliang},
title = {A dynamic task partitioning method for CPU–FPGA heterogeneous systems for computing intensive workloads},
journal = {Journal of Vibration and Control},
year = {2026},
doi = {10.1177/10775463261508672}
}
EUVIMED