Performance Analysis of Big Data ETL Process over CPU-GPU Heterogeneous Architectures

While GPUs have been utilized in the analysis stage of big data processing, the demand for GPU use in the extract-transform-load (ETL) stage has recently been increasing. There have been several research efforts to use GPUs for query processing in database systems. However, most of them are mainly focused on batching input data as much as possible to increase the overall throughput with GPUs. Moreover, they try to execute a query plan by only using a single device (only CPU or only GPU), based on the assumption that the PCIe transfer overhead between CPU and GPU is significant. In contrast, this paper presents several interesting observations and reveals that it is better to use a CPU-GPU heterogeneous query plan in some cases. For example, the heterogeneous query plan proved to be efficient when the input data size is small. This is because the PCIe overhead is not as serious as expected for small-sized data and can be offset with performance gains by partially using the GPU. In addition, even when the data size is large, it is sometimes better to use the CPU-GPU heterogeneous query plan rather than only the GPU plan, since some relational operations prefer the CPU over the GPU. To demonstrate these uses, this paper presents several experimental results and their analyses with Spark SQL in a CPU-GPU heterogeneous computing environment.