Remote Submission and Parallel Computation of Compute

advertisement
Remote Submission and Parallel Computation of
Compute-intensive Applications
Ye Yang and Roy P. Pargas
Department of Computer Science, Clemson University
Clemson, SC 29634
{yangy, pargas}@cs.clemson.edu
Abstract
This paper describes a software system currently being developed at Clemson University
in which a client provides recently obtained data to a remote server running a computeintensive algorithm. The results of the algorithm provide information necessary for the
client to continue his or her activities. To improve performance and speed up delivery of
the results, the server distributes the data among multiple processors and assembles
partial output from each processor into a coherent whole before sending the final results
to the client.
The specific application presented in this paper is a fluorescent lifetime image
reconstruction system for breast cancer detection. It involves a doctor or inspector at a
hospital (the client) performing diagnostic tests on a patient suspected of breast cancer.
An experimental instrument optically scans the patient’s breast and generates a file of
data which is sent to the server via the web. The data is processed by a numerical
algorithm running on a server at Clemson. The algorithm reconstructs the image of the
breast in parallel, highlighting potential tumors. The resulting image is sent back to the
doctor for analysis.
The numerical process is based on a set of coupled diffusion equations which are used to
describe the propagation of excitation and fluorescent emission light in multiply
scattering media (such as a breast). A finite element based reconstruction algorithm
combined with Marquardt and Tikhonov regularization methods are used to obtain the
fluorescence lifetime images.
This paper briefly describes the numerical algorithm but focuses primarily on the
software system which uses Java servlets to collect the data from the client and remote
method invocation (Java RMI) to distribute the data to multiple processors. The entire
software system was written in Java. The output of the numerical algorithm, combined
with the corresponding finite element mesh information, are input into a mathematical
software package called Matlab which is used to produce the final images. We will report
on the results of simulation runs as well as the speedup obtained.
Future work includes a refinement of the algorithm to incorporate adaptive mesh
techniques. The expectation is that such techniques will improve the accuracy of the
reconstructed images.
Download