Ubiquitous FPGA Access for Data Intensive Computing Julio Dondo, Francisco Sanchez Molina, Fernando Rincon, Francisco Moya, Juan Carlos Lopez University of Castilla-La Mancha - UCLM Ciudad Real, Spain Email: juliodaniel.dondo, francisco.smolina, fernando.rincon, francisco.moya, juancarlos.lopez @uclm.es Abstract—The availability of ubiquitous distributed reconfig- urable devices allows the implementation of different function- alities everywhere using remote resources. This work presents a development model introducing an efficient and secure manner of managing these remote and distributed reconfigurable resources. This model is a promising model to be used in scenarios where data intensive and high performance computing are needed, because it allows acceleration of applications using distributed hardware. Index Terms—reconfigurable logic; distributed architectures; FPGA; HPC; I. I NTRODUCTION A. Data Intensive Computing Data-intensive computing capabilities are fundamental for data-intensive scientific research such as biological systems, weather prediction, simulation of statistical mechanics, as well as analyses of huge volumes of different data types to and from several distributed sources. Examples of these situations can be found in power grids (Smart Grid), or in a distributed-sensors network (cameras, sensors, etc.) used for security, and disaster prevention, among others. These networks may acquire large amounts of data and the data processing also may need to be done locally, and in real time, requiring some computation- intensive task, e.g. edge detection in images. Even more, in these networks each node can be power-constrained and then a data-intensive task can be limited by power consumption. In a wireless sensor network, nodes are low-cost sen- sors operating in an environment with limited processing power and restricted battery autonomy. In this scenario data processing can be done locally, using nodes equipped with dedicated processors. Only the results of the processed data are transmitted to the base station. This approach has the advantage of consuming less bandwidth when sending data, but it has to deal with power restrictions. Another approach is to process data remotely. In this case a high volume of data is transferred through the network. One way to solve the disadvantages of both approaches is through the use of low power reconfigurable hardware (FPGAs) dedicated to per- forming data processing locally. Data Processing algorithms can be implemented in these FPGAs obtaining n improved time response with low power consumption. In a system with several sensors and actuators, FPGAs can be configured in order to perform different data-intensive processing, liberating the sensors from the data processing [7] B. High Performance Computing Currently there exists a great demand for high performance computing due largely to the requirements of scientific re- search. For these applications, the use of accelerators offers a qualitative improvement in terms of computing power and consumption, compared with the classical solutions of general purpose computing. FPGA-based scientific computation is one of the solutions in the scientific community to improve the response time for numerically-intensive computation. FPGA-based systems are faster than a software-only approach in terms of computational power [5] [1]. Approaches for high performance reconfigurable comput- ing, (HPRC), integrates both processors and FPGAs into a parallel architecture. HPRC can achieve an improvement in speed, size, and cost of several orders of magnitude over conventional supercomputers [6]. Another solution widely used to improve response time in intensive computation consists of the shared use of dis- tributed resources [8]. Grid computing allows the connection of scattered computational resources enabling the use of their computer capacity, data, and storage space, regardless of their location. To deal with such diversity of resources and local administration policies virtualization technology has been used to abstract the demands and use of resources from the sujacent hardware platform [3]. Facilities to accelerate computational problem using recon- figurable hardware, and the availability of these reconfigurable resources as in a distributed grid system, can be combined to create a computing resource capable of making significant breakthroughs in data-intensive computing. Distributed reconfigurable resources can be integrated ac- cording to the grid model, allowing the creation of a distributed reconfigurable platform for data-intensive computing as in sensor network or Smart Grids. Even more, this distributed reconfigurable platform is ideal also for high performance computing for the resolution of scientific computational prob- lems. This platform is depicted in figure 1. This platform that we call Reconfigurable Grid (R-GRID) is not just a set of spatially distributed FPGAs but also includes a set of functionalities to provide the ability of a transparent implementation of concurrent distributed applications, in order to obtain maximum benefits from reconfigurability and spatial distribution of resources. R-GRID platform was designed to