In-memory analytical query processing on GPUs

Abstract

The high global memory bandwidth and the large number of parallel cores available in modern Graphics Processing Units (GPUs) make them ideal for highperformance Online Analytical Processing (OLAP) systems. However, it is challenging to design efficient high-performance query processing systems for GPUs due to the following reasons: 1) the rapid evolution of GPU hardware in the last decade, 2) the significant differences in the hardware architecture of GPUs when compared to CPUs, 3) the high overhead of moving the data between the CPU and GPU and 4) the small global memory size of a single GPU that necessitates the access of remote data over PCIe or NVLink interconnects when processing large data sets. In this thesis, we study existing query processing systems for GPUs and propose techniques to improve query execution efficiency on GPUs.We begin by studying the performance of hash join, which is one of the most compute and memory intensive relational operator in OLAP systems. Specifically, we first revisit the major GPU hash join implementations in the past decade that were designed to execute on a single GPU. We then detail how these implementations take advantage of different GPU architecture features and conduct a comprehensive evaluation of their performance and cost-efficiency using different generations of GPUs. This helps shed light on the impact of different GPU architecture features on the performance of the hash join operation and identify the factors guiding the choice of these features. We further study how data characteristics like skew and match rate impact the performance of GPU hash join implementations. Novel techniques to improve the performance of the hash join operation when joining input relations with severe skew or high match rate were also proposed as part of this study. Our evaluation finds that the proposed techniques help avoid any performance degradation when joining skewed input relations and achieve up to 2.5x better performance when joining input relations with high match rate.Next, we extend our study on the hash join operation to modern multi-GPU architectures. The recent scale-up of GPU hardware through the integration of multiple xi Contentsxii GPUs into a single machine and the introduction of higher bandwidth interconnects like NVLink 2.0 has enabled new opportunities of relational query processing on multiple GPUs. However, due to the unique characteristics of GPUs and the interconnects, existing hash join implementations spend up to 66% of their execution time moving the data between the GPUs and achieve lower than 50% utilization of the newer high bandwidth interconnects. This leads to extremely poor scalablity of hash join performance on multiple GPUs, which can be slower than the performance on a single GPU. In this thesis, we propose MG-Join, a scalable partitioned hash join implementation on multiple GPUs of a single machine. In order to effectively improve the bandwidth utilization, we develop a novel multi-hop routing for cross-GPU communication that adaptively chooses...

pLUTo: Enabling Massively Parallel Computation in DRAM via Lookup Tables

“…Processing-in-Memory (PiM) is a promising paradigm that augments a system's memory with compute capability [1][2][3][4][5] to alleviate the data movement bottleneck between processing and memory units [2,[6][7][8][9][10][11][12][13][14][15][16][17]. PiM architectures can be classified into two categories [1,18]: 1) Processing-near-Memory (PnM), where computation takes place in dedicated processing elements (e.g., accelerators [11,, processing cores [11, 30-32, 41, 44-58], reconfigurable logic [59][60][61][62][63]) placed near the memory array (e.g., [11,), and 2) Processingusing-Memory (PuM), where computation takes place inside the memory array by exploiting intrinsic analog operational properties of the memory device (e.g., [41,55,).…”