Distributed Processing of High-speed Stereo Particle Image Velocimetry Data with a Low Power Beowulf Cluster

In this paper, we present our design of a portable cluster supercomputer created specifically for processing PIV image data. To make the system portable for a laboratory environment, the cluster computer is required to run on minimal power and be light weight. The final cluster design was based on the Intel R ©QuadCore Xeon R ©X3210 low-power processors. This hardware setup allows for a total system consisting of twelve processing nodes with a total of 48 CPU cores, one master node with 5 TB of storage, and a Gigabit Ethernet connection. The total configuration along with the case has an approximate gross weight of 390 lbs and works well on one standard 120-V, 20-Amp electric circuit. With this cluster computer, a speedup of about 28 relative to standard serial processing of a PIV data set could be achieved. ∗Address all correspondence to this author. Introduction Time-Resolved Stereo Particle Image Velocimetry (TRSPIV) is a revolutionary measurement technique with the potential to fundamentally alter the way we approach fluid measurements. Recent results [?] demonstrate that TRSPIV can reveal the full 3-D structure of a flow, which was previously only possible with Direct Numerical Simulation. However, the current state-of-the-art for processing these data sets falls far short of the state-of-the-art in high-speed data acquisition. TRSPIV evolved from the original Particle Image Velocimetry (PIV) approach. The PIV process generates a planar 2component velocity vector field at an instant in time. Stereo Particle Image Velocimetry is performed by adding a second camera to a standard PIV system. The two cameras observe the laser sheet from the sides at angles allowing them to sense flow perpendicular to the sheet in addition to flow in the plane of the laser sheet. As a result, all three components of velocity can be computed in the plane of the laser sheet. TRSPIV is possible when the images can be acquired sufficiently-fast to resolve the 1 Copyright c © 2007 by ASME Figure 1. The initial system design for the TRSPIV system. The TRSPIV system acquisition computer is linked directly to the the in-lab Cluster. time-scales of interest. Recent advances in digital photography and solid-state lasers make it possible to acquire images at up to 3000 frames per second. However, as the ability to acquire large samples very quickly has been realized, processing speed has not kept pace. Our present (and typical) 2-D PIV acquisition computer at USU, purchased recently, would require over five hours to process the data that can be acquired in one second with the TRSPIV system. To decrease the computational time, parallel processing has been implemented using a Beowulf cluster [?]. Beowulf cluster supercomputers are built from commodity parts and provide low cost parallel processing power. At USU we have developed a low-power Beowulf cluster integrated with the data acquisition system of a TRSPIV system with a low power cluster supercomputer specifically designed for the data processing of large experimental image data sets. This approach of integrating the PIV system and the Beowulf cluster eliminates the communication time, thus speeding up the process. In addition to improving the practicality of TRSPIV, this system will also be useful to researchers performing any PIV measurement where a large number of samples are required. This paper will describe the hardware and software implementation of our approach.