Tiny GPU Cluster for Big Spatial Data: A Preliminary Performance Evaluation

GPU-equipped computing nodes have much higher ratios between floating point computing power (in the order of TFlops and is fast growing) and network bandwidth (in the order of Gbps and remains stable) than regular computing nodes at which Hadoop-based systems are targeting. The gap makes efficient and scalable processing of large-scale data challenging, especially for geo-referenced spatial (or geospatial) data, whose processing is both data intensive and computing intensive. We aim at developing a tiny GPU cluster using Nvidia Tegra K1 (TK1) System on Chip (SoC) boards as a downscaled, low-cost GPU cluster for Big (Spatial) Data research. The tiny GPU cluster is equipped with standard gigabyte Ethernet network while has much less computing power and energy footprint when compared with a regular GPU cluster and represents a new platform with more balanced compute to communication ratio. We have ported our implementations of both single-node technologies for point-in-polygon test based spatial joins and the lightweight distributed execution engine originally developed for regular clusters to the tiny GPU cluster. We evaluate its performance on two real world geospatial applications with various settings and experiment results have demonstrated good scalability. Preliminary analysis on the scaling effect between the tiny cluster and a regular Amazon EC2 cluster using a simplified model suggest that the ARM-based CPU of the TK1 board is likely to achieve better energy efficiency while the Nvidia GPU of the TK1 board might be less efficient when compared with desktop/server grade GPUs, in both standalone and 4-node cluster settings.