Parallel Computing Sub-System

The Parallel Computing Hardware

The Parallel Computing Sub-System is the “performance provider” of SMART-PIV. It consists of two major modules:

1. The Parallel Computing Hardware
2. The Parallel Computing SW

The Parallel Computing Hardware (Figure 1) aims at providing a high-performance, scalable platform for the parallel evaluation of acquired images and the parallel numerical analysis simulation. In order to achieve this, the Parallel Computing HW is built upon a new philosophy based on the use of commodity components to build high-performance and cost-efficient parallel computing platforms. This new approach assembles commodity PCs and interconnects them with the use of a widely used interconnection network to implement a so called PC Cluster.

Figure 1: Photo of SMART-PIV Parallel Computing HW

PC Clusters are an increasingly popular parallel computing platform, having already dominated the small and medium-scale parallel applications and constantly penetrating the large-scale ones. In November 2002, it was reported that 18% of the 500 most powerful supercomputers worldwide were PC clusters, while in June 2003 this percentage had grown to 29.8%.

There are two key features of PC Clusters, responsible for their success and popularity:

Commodity HW enables for cost-effective, up-to-date systems to be built that are easily scalable and provide high computational power. Note that using the most up-to-date processors one can always deliver the highest possible performance to the parallel applications, whereas this is not the case for vendor-built parallel machines. When a new parallel architecture project is launched by a manufacturer, the latest microprocessor is used to design and implement the configuration. However, at the time the architecture reaches the market (at least six months later) this CPU has become rather obsolete.

Open-Source SW extensively used in computer clusters (Linux OS, GCC compiler, MPI parallel library), keeps the installation cost low and provides ease of maintenance.

PC Clusters provide the highest performance to cost ratio, and are consequently the most popular solution for small and medium-scale parallel applications. As far as the computational needs of the SMART-PIV application are concerned (image processing and numerical analysis), it is believed that a Linux Cluster is the most suitable parallel architecture. Built upon the latest processor technology, the SMART-PIV cluster is expected to deliver the required performance. Due to its scalability, more processing nodes can be added at any time in the future, further contributing to performance if necessary. It is also a fact that, at the same cost, no other supercomputing infrastructure can compete with the performance and scalability offered by a PC Cluster. It seems that PC Clusters are the unique affordable solution for similar small to medium-scale applications, similar to the image processing and numerical analysis for SMART-PIV.

The Parallel Computing HW has already been implemented. The basic components of the platform are:

4 Processing nodes:

1 File Server:

Interconnection Network:

The general design is shown in Figure 2. The interconnection network between the processing nodes is Fast Ethernet. The Cluster of the processing nodes is connected to the File Server with a Gigabit Ethernet link. The link between the File Server and the Recording PC is Gigabit Ethernet as well.

A typical operational scenario would require the camera attached at the Recording PC to record images at a rate of 15-60 Mbytes/sec. These images are transferred directly to the File Server where they are made available for the cluster processing nodes. The nodes read the images from the File.

Server perform image processing and evaluation write the results (vector fields) back to the File Server for further processing and/or visualization.

Figure 2: Design of Parallel Computing HW