Shisha: Online Scheduling of CNN Pipelines on Heterogeneous Architectures

Paper in proceeding, 2023

Many modern multicore processors integrate asymmetric core clusters. With the trend towards Multi-Chip-Modules (MCMs) and interposer-based packaging technologies, platforms will feature heterogeneity at the level of cores, memory subsystem and the interconnect. Due to their potential high memory throughput and energy efficient core modules, these platforms are prominent targets for emerging machine learning applications, such as Convolutional Neural Networks (CNNs). To exploit and adapt to the diversity of modern heterogeneous chips, CNNs need to be quickly optimized in terms of scheduling and workload distribution among computing resources. To address this we propose Shisha, an online approach to generate and schedule parallel CNN pipelines on heterogeneous MCM-based architectures. Shisha targets heterogeneity in compute performance and memory bandwidth and tunes the pipeline schedule through a fast online exploration technique. We compare Shisha with Simulated Annealing, Hill Climbing and Pipe-Search. On average, the convergence time is improved by ∼ 35 × in Shisha compared to other exploration algorithms. Despite the quick exploration, Shisha’s solution is often better than that of other heuristic exploration algorithms.