Core architecture-dependent race conditions in heterogeneous multicore processors

Abstract

Heterogeneous big.LITTLE SoCs are central to edge AI but complicate parallel software reliability. We study race conditions in an OpenMP FSRCNN model and a histogram workload on an Orange Pi 5. The Lost Update Ratio reverses with workload: Cortex-A76 loses fewer updates than Cortex-A55 on FSRCNN and more on histogramming. A factorial experiment at fixed target space attributes this not to address availability, which out-of-order cores speculate over, but to the collision structure of the access pattern, which changes the ratio by −50% on A55 and two orders of magnitude on A76. Correctness is not portable across core types.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint