.png){kind=icon}
.png){kind=icon}
Return
A portable framework with generalized runtime features for task graph execution and concurrent multi-application deployment on heterogeneous systems
.png){kind=icon}
PRE
AI
DOI：10.1016/j.future.2025.108184.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Heterogeneous computing platforms are widely adopted to meet the diverse performance demands of modern applications, but they present a key challenge of balancing programmability with performance portability. This work introduces a unified and portable framework that addresses this trade-off by integrating the CEDR runtime system with the Taskflow task-graph programming model. By combining CEDR’s dynamic scheduling with Taskflow’s parallelization capabilities, our system simplifies application development while enabling efficient execution across CPUs, GPUs, and FPGAs. We demonstrate its portability on a broad range of heterogeneous SoC and HPC-scale systems. The framework supports concurrent execution of multiple applications through a centralized coordination layer, overcoming limitations of isolated execution contexts and achieving up to 1.23 × speedup on highly heterogeneous platforms. Our solution further generalizes runtime features such as streaming input handling and cached scheduling across applications, yielding up to 6.08 × improvements in execution time and a 29.60 × reduction in scheduling overhead. Experimental evaluations demonstrate consistent improvements in execution time, validating the effectiveness of this integrated and extensible design.
Journal
F
IF:
0
Papers:
642
Citations:
0