Cloud computing is a growing computing paradigm that enables important computing capabilities including large-scale data processing. Cloud allows users to leverage heterogeneous hardware for high-performance computation and memory without the capital expenditure to acquire the hardware devices. In particular, Graphics Processing Units (GPUs) are computational devices that achieve much higher computational power through massive parallelism and have been widely used in applications including graphics, machine learning, and scientific computing. In this project, we use GPUs to accelerate data analytics applications to achieve much higher processing speed. Specifically, we conducted the following two tasks:

1. Data compression in GPU. 

A key constraint of existing GPU-based data analytics systems is the limited memory capacity in GPU devices. So that only small workloads can leverage GPU acceleration. Data compression is a powerful technique that can mitigate the capacity limitation in two ways: (1) fitting more data into GPU memory and (2) speeding up data transfer between CPU and GPU. However, compression schemes for GPU today are still limited in compression ratio and/or decompression speed. We identify two limiting factors of existing approaches. First, existing decompression solutions require multiple passes of scanning the global memory to decode layers of compression schemes, incurring significant memory traffic and hurting performance. We present the tile-based decompression model to decompress encoded data in a single pass over global memory and inline with query execution. Second, we develop an efficient implementation of bit-packing-based compression schemes and their optimization techniques in the context of GPU. Our evaluation shows that our schemes can achieve similar compression rates to the best state-of-the-art compression schemes in GPU (i.e., nvCOMP) while being 2.2x and 2.6x faster in decompression speed and query running time.

2. Hybrid CPU/GPU processing. 

Heterogeneous CPU-GPU query execution is another compelling approach to mitigate the limited GPU memory capacity and PCIe bandwidth between CPU and GPU. However, the design space of heterogeneous CPU-GPU query execution has not been fully explored. We improve state-of-the-art CPU-GPU data analytics engine by optimizing data placement and heterogeneous query execution. First, we introduce a semantic-aware fine-grained caching policy which takes into account various aspects of the workload such as query semantics, data correlation, and query frequency when determining data placement between CPU and GPU. Second, we introduce a heterogeneous query executor which can fully exploit data in both CPU and GPU and coordinate query execution at a fine granularity. We integrate both solutions in a novel hybrid CPU-GPU data analytics engine that we developed. Evaluation on the Star Schema Benchmark shows that the semantic-aware caching policy can outperform the best traditional caching policy by up to 3x.

Last Modified: 01/23/2023
Modified by: Xiangyao Yu