Heterogeneous memory systems that comprise memory nodes with disparate architectural characteristics (e.g., DRAM and high-bandwidth memory (HBM)) have surfaced as a promising solution in a variety of computing domains ranging from embedded to high-performance computing. Since deep learning (DL) is one of the most widely-used workloads in various computing domains, it is crucial to explore efficient memory management techniques for DL applications that execute on heterogeneous memory systems. Despite extensive prior works on system software and architectural support for efficient DL, it still remains unexplored to investigate heterogeneity-aware memory management techniques for high-performance DL on heterogeneous memory systems. To bridge this gap, we analyze the characteristics of representative DL workloads on a real heterogeneous memory system. Guided by the characterization results, we propose HALO, hotness- and lifetime-aware data placement and migration for high-performance DL on heterogeneous memory systems. Through quantitative evaluation, we demonstrate the effectiveness of HALO in that it significantly outperforms various memory management policies (e.g., 28.2 percent higher performance than the HBM-Preferred policy) supported by the underlying system software and hardware, achieves the performance comparable to the ideal case with infinite HBM, incurs small performance overheads, and delivers high performance across a wide range of application working-set sizes.

Heterogeneous embedded systems have surfaced as a promising solution for accurate and efficient deep-learning inference on mobile devices. Despite extensive prior works, it still remains unexplored to investigate the system-software support that efficiently executes inference workloads by judiciously considering their performance and energy heterogeneity, communication overheads, and constraints. To bridge this gap, we propose MOSAIC, heterogeneity-, communication-, and constraint-aware model slicing and execution for accurate and efficient inference on heterogeneous embedded systems. MOSAIC generates the efficient model slicing and execution plan for the target inference workload through dynamic programming. MOSAIC significantly reduces inference latency and energy, exhibits high estimation accuracy, and incurs small overheads.

Memory bandwidth is a highly performance-critical shared resource on modern computer systems. To prevent the contention on memory bandwidth among the collocated workloads, prior works have investigated memory bandwidth partitioning techniques. Despite the extensive prior works, it still remains unexplored to characterize the widely-used memory bandwidth partitioning techniques based on various metrics and investigate a hybrid technique that employs multiple memory bandwidth partitioning techniques to improve the overall efficiency.