Title: Rapid IO-based space system architectures for synthetic aperture radar and ground moving target indicator
CITATION PDF VIEWER THUMBNAILS PAGE IMAGE ZOOMABLE
Full Citation
STANDARD VIEW MARC VIEW
Permanent Link: http://ufdc.ufl.edu/UF00094732/00001
 Material Information
Title: Rapid IO-based space system architectures for synthetic aperture radar and ground moving target indicator
Physical Description: Book
Language: English
Creator: Bueno, David
Conger, Chris
Leko, Adam
Troxel, Ian
George, Alan D.
Publisher: Bueno et al.
Place of Publication: Gainesville, Fla.
Copyright Date: 2005
 Record Information
Bibliographic ID: UF00094732
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.

Downloads

This item has the following downloads:

HPEC05_UF ( PDF )


Full Text



RapidlO-based Space System Architectures for
Synthetic Aperture Radar and Ground Moving Target Indicator

David Bueno, Chris Conger, Adam Leko, Ian Troxel, and Alan D. George
{bueno, conger, leko, troxel, george}@hcs.ufl.edu
High-performance Computing and Simulation (HCS) Research Laboratory
Department of Electrical and Computer Engineering, University of Florida


Abstract
The design of space systems capable of performing real-
time Synthetic Aperture Radar (SAR) is a ;.-; r...,, it
challenge in HPEC due to the high processor, memory,
and network requirements imposed by SAR. However,
building a system to support SAR and other Space-Based
Radar (SBR) ,uig. ,riim, simultaneously is an even greater
challenge. This presentation describes simulation results
from a SAR application i. .... rii in parallel on an
embedded multiprocessor satellite system equipped with a
RapidlO interconnection network. We consider several
variations on RapidlO network parameters as well as
parallel decompositions of the SBR ,uig. 'drimi, to design a
system to support SAR and Ground Moving Target
Indicator (GMTI) simultaneously. In addition, we describe
our RapidlO testbed that is used to validate our network
models and present results d.,. ma ,,tai ,,, the accuracy of
our network models under varying traffic conditions.

Introduction
The large dataset sizes and real-time requirements of SBR
applications demand a high-performance interconnect
architecture capable of meeting throughput requirements
on the order of several gigabits per second. In previous
work [1], we presented simulation results for systems
performing GMTI in parallel on system architectures built
around the RapidIO embedded systems interconnect. In
this presentation, we significantly extend our previous
work with new simulations of a parallel SAR application
executing on a RapidIO-based system. RapidIO is an
emerging open standard for high-speed, embedded packet-
switched interconnection networks which supports data
rates up to approximately 60 Gbps. RapidIO is the latest
commercial-off-the-shelf (COTS) technology to be
considered practical for inclusion in military embedded
networks to improve cost-effectiveness and scalability.
Moving from bus designs to switched interconnects
substantially increases the cost-effectiveness, robustness,
and network performance of future embedded systems.
SAR has recently gained popularity for applications
such as remote sensing, surveillance, and target
recognition [2]. SAR contains some high-level similarities
to GMTI, such as the ability to break processing into
stages and the repartitioning of data between different
stages [3-4]. However, there are also distinct and
important differences, which provide an interesting
comparison of system performance and limitations for
systems built to support both GMTI and SAR. In addition
to significant computational changes, examples of major
differences of SAR vs. GMTI include much larger memory


requirements and therefore associated adjustments in data
movement strategies, as well as relaxed time constraints,
allowing significantly more time for processing of SAR
images relative to GMTI data cubes. For both SAR and
GMTI, the processing deadline for a single dataset is
referred to as a Coherent Processing Interval (CPI). While
our GMTI algorithm mandates a 256 ms CPI, SAR allows
up to 16 seconds per CPI. To provide a more fair and
thorough comparison of system performance for the
different SBR algorithms, we also present results for a new
parallel partitioning for GMTI which more closely matches
the strategy of our SAR algorithm, which complements the
three GMTI partitionings presented in [1] and [5].
Computer-based simulation provides a cost-effective
tool for evaluating future satellite systems for SBR and
other payload-processing applications. Our discrete-event
simulation models are built and simulated using the
commercial simulator MLDesigner from ML Design
Technologies. Our virtual prototyping environment for
RapidIO systems incorporates moderate-fidelity systems
and components including RapidIO switches, end-points,
and processor models. To ensure that our simulation
results are accurate and instill confidence in our results, we
have created a RapidIO testbed using Xilinx RapidIO
cores, FPGAs, and development boards.

RapidlO Testbed for Model Validation
Our RapidIO testbed currently consists of two RapidIO
endpoints connected via an 8-bit parallel LDVS, 250 MHz
RapidIO interface. The RapidIO endpoints are
implemented on Xilinx development boards each
containing a Virtex-II Pro FPGA, with plans for expansion
to include one or more Tundra 4-port, 250 MHz RapidIO
switches. We measure RapidIO throughput and packet
latencies under varying traffic conditions and calibrate our
models to closely match the performance of the RapidIO
testbed. Due to space and time limitations, results are not
included here. The formal presentation will contain a
variety of validation results over a range of RapidIO
packet types, sizes, and traffic conditions.

Experimental Setup
Our baseline SAR algorithm processes a 2D data image of
size 2 GB, where each data element is a 64-bit complex
integer. The high memory requirements of SAR images
dictate a different parallelization approach from the three
methods used for partitioning of GMTI in [1]. Instead of
directly spreading out the entire image over the system
nodes in some fashion, SAR requires the entire image be
stored in and processed out of a global memory board.







The data image is then inputted in "chunks" by the
processing nodes, where each then processes a chunk,
writes it back to global memory, reads in another chunk,
and repeats. Rather than explicit "corner turns" to
redistribute data between stages, data redistribution occurs
implicitly as nodes read in data from global memory
partitioned along the appropriate dimension for the stage
whose chunks are being processed. The disadvantage of
this approach is that redundant data communication
frequently occurs, as each data item must be written to
global memory after each stage and then read in again
during the subsequent stage.
A complete description of all system designs and
network tradeoffs performed in the course of this study
will appear in the full presentation. Due to space
limitations here, a condensed version follows. RapidIO
system parameters are similar to those used in [1] and [5].
Our baseline system backplane is a composed of four 8-
port RapidIO switches and provides non-blocking
connections for up to eight boards (including a global
memory board), with four dedicated RapidIO links
between the backplane and each board. Each processor
board has an 8-port switch and four compute nodes.

SAR: Global-Memory, Chunk-Based

14 5
14 -*-Double-Buffered
S-m- Synchronized
S13 5 Unsynchronized
13
E Z/
S125
o 12


11
11 s ------------------------
64KB 128KB 256KB 512KB 1MB 2MB 4MB
System-Level Chunk Size
Figure 1- Variants of Parallel SAR Algorithm


Results

Figure 1 shows a summary of CPI latency results for three
different variants of the parallel SAR algorithm on this
multiprocessor platform. For the unsynchronized case,
network contention severely hampers performance as
chunk size increases. This contention is largely due to
multiple nodes requesting transactions from the same
global memory port at the same time. The synchronized
case uses a token to only allow one node to attempt to read
from global memory at a time, creating more efficient
network usage and improving performance, especially for
large chunk sizes. The double-buffered case assumes that
nodes possess sufficient memory to store and process a
chunk in memory while receiving the subsequent chunk.
Double buffering achieves the best performance for small
chunk sizes, but introduces even higher levels of
contention for larger chunk sizes. A broad array of results
describing additional algorithm approaches and system


design tradeoffs will be included in the final presentation,
including SAR vs. GMTI, but are omitted in this abstract.

Conclusions
The incorporation of RapidIO in future satellite payload
processing systems is likely to improve performance as
well as cost-effectiveness of embedded SBR platforms.
This new work builds off of our previous work in
modeling and simulation of RapidIO and GMTI,
demonstrating that effective, real-time SAR processing can
be performed over a RapidIO-based network with HPEC
onboard a satellite. Results show the importance of
intelligently synchronizing communications over the
RapidIO network and demonstrate significant performance
gains obtainable by double-buffering of data cube chunks.
Our work shows that RapidIO-based switched interconnect
designs have the potential to far outperform traditional bus
designs in embedded systems, allowing real-time GMTI
and SAR processing to occur on-board a satellite. Future
directions for this work will include the study of fault-
tolerant RapidIO-based space system architectures for
SBR and other payload-processing applications. Future
experimental and simulative work will also feature
hardware-reconfigurable computing devices (e.g. FPGAs)
working alongside general-purpose processors and other
computing devices connected via a RapidIO fabric.

Acknowledgements

We wish to thank Honeywell Defense and Space
Electronic Systems (DSES) in Clearwater, FL for support
of this research. We also extend thanks to Xilinx for their
generous donation of hardware and IP cores, as well as
MLDesign Tech. for donation of simulation software.

References

1. D. Bueno, C. Conger, A. Leko, I. Troxel, and A.
George, "Virtual Prototyping and Performance
Analysis of RapidIO-based System Architectures for
Space-Based Radar," Proc. High-Performance
Embedded C. ,,wa,,' (HPEC) Workshop, MIT
Lincoln Lab, Lexington, MA, Sep. 28-30, 2004.
2. C. Cho, "Performance Analysis for Synthetic Aperture
Radar Target Classification," MSEE Thesis, Mass.
Institute of Technology, Feb. 2001.
3. P. Meisl, M. Ito, and I. Cumming, "Parallel Synthetic
Aperture Radar Processing on Workstation
Networks," Proc. 10th International Parallel
Processing Symp., Apr. 15-29, 1996, pp. 716-723.
4. S. Plimpton, G. Mastin, and D. Ghiglia, "Synthetic
Aperture Radar Image Processing on Parallel
Supercomputers," Proc. 1991 ACM/IEEE Conf on
Supercomp., Albuquerque, NM, 1991, pp. 446-452.
5. D. Bueno, A. Leko, C. Conger, I. Troxel, and A.
George, SiImi.l\ c Analysis of the RapidIO
Embedded Interconnect Architecture for Real-Time,
Network-Intensive Applications," Proc. 29th IEEE
Conf on Local Computer Networks (LCN) via the
IEEE Workshop on High-Speed Local Networks
(HSLN), Tampa, FL, Nov. 16-18, 2004.




University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs