Massively parallel skyline computation for processing-in-memory architectures

摘要: Processing-In-Memory (PIM) is an increasingly popular architecture aimed at addressing the 'memory wall' crisis by prioritizing integration of processors within DRAM. It promotes low data access latency, high bandwidth, massive parallelism, and power consumption. The skyline operator a known primitive used to identify those multi-dimensional points offering optimal trade-offs given dataset. For large multidimensional dataset, calculating extensively compute intensive. Although, PIM systems present opportunities mitigate this cost, their execution model relies on all operating in isolation with minimal exchange. This prohibits direct application optimizations which are inherently sequential, creating dependencies intermediate results that limit maximum throughput, require expensive merging phase. In work, we address these challenges introducing first algorithm for architectures, called DSky. designed be massively parallel throughput efficient leveraging novel work assignment strategy emphasizes load balancing. Our experiments demonstrate it outperforms state-of-the-art algorithms CPUs GPUs, most cases. DSky achieves 2× 14× higher compared solutions competing CPU GPU architectures. Furthermore, showcase DSky's good scaling properties intertwined PIM's ability allocate resources added cost. addition, order magnitude better energy consumption GPUs.