Automated Interface Generation for Remote Access to Adaptive

                                  Computing Resources


Christopher C. Doss and Clay Gloster
NC State University

Please send all correspondence to Clay Gloster 
Clay Gloster, Jr., Ph.D., P.E.
NC State University
Department of Electrical & Computer Engineering
Box 7914
Raleigh, NC 27695-7914
email:     gloster@eos.ncsu.edu
telephone: (919) 515-7348	
fax:       (919) 515-2285

Adaptive computing has emerged as a viable computing option for implementing 
computationally intensive applications. Herein, an adaptive computer is a 
computer system that contains a host processor attached to a programmable logic 
device, i.e. a field programmable gate array (FPGA). These computing systems 
combine the flexibility of general purpose processors with the speed of 
application specific processors. By mapping hardware to FPGAs, the system 
designer can cater the hardware to a specific application. Because the arrays do 
not have a fixed structure, they can be reconfigured for other applications.

Due to the excessive cost of FPGAs and the limited availability of engineers 
with extensive hardware design expertise, research organizations often have 
access to few adaptive computers. This results in a limited number of 
applications that are ultimately mapped to this environment. Our research 
extends the use of an adaptive computer to act as an application server, 
allowing client computers to attach to it and utilize its available resources. 
This allows users to benefit from the potential performance gains that are 
offered by an adaptive computer without the need to gain substantial expertise 
in the field.

This paper presents the initial phase of the RARE (Remote Adaptive computing 
Resource Environment) project.   RARE provides users access to adaptive 
computing resources from a remote site. The initial phase of the project 
involves the creation and verification of FEW (FilE Writer), a tool that 
examines an input specification of an adaptive computing resource module and 
automatically generates all files required to interface a remote client to the 
application server that implements this module.

Written in Java and C, RARE allows clients to access the field programmable gate 
arrays of an adaptive computer. Host to FPGA interface modules are written in C, 
while the client to server interfaces are written in Java. The goal of the work 
presented in this paper is to automatically generate these interfaces, relieving 
the user of this task. The adaptive compute server manages a set of routines 
that have been previously implemented by hardware design engineers. All 
potential users of the resource obtain the list of routines and how to utilize 
them from the RARE web site. 

We prove the viability of RARE by implementing a multispectral image 
classification algorithm to the FPGA coprocessor. This algorithm, based on a 
probabilistic neural network (PNN), classifies image pixels, originating from a 
satellite, as belonging to certain classes (water, forest, tundra, etc.). A 
client processor is able to send pixel data over the Internet to the adaptive 
computer server. The server performs the calculations and sends the results to 
the client. The client and server communicate with each other using Java Remote 
Method Invocation (RMI) and the server communicates with the FPGA coprocessor 
using the Java Native Interface (JNI). We then developed an input specification 
of the PNN algorithm for use with the program FEW. FEW automatically generates 
all required client and server RMI file stubs, as well as the server JNI file 
stubs.  Results indicate that the use of the FPGA coprocessor provides an order 
of magnitude speedup over the implementation executed on the host alone.