Parallelization of a Non-Linear Analysis Code

1
Parallelization of a Non-Linear Analysis Code
Travis Whitlow Alabama A M University Research
Alliance in Math and Science Computer Science and
Engineering Division Mentors Dr. Lee Hively and
Dr. Jim Nutaro http//www.csm.ornl.gov/Internship
s/rams_07/abstracts/t_whitlow.pdf
Research Over the last decade, ORNL has developed
and patented a novel approach for forewarning of
a large variety of machine and biomedical events.
One of the most amazing uses of this research
involves accurate prediction of epileptic
seizures with up to 4.5 hours of forewarning. The
analysis involves 27 GB of brain wave data with
40 seizure datasets and 20 non-event datasets.
The majority of this analysis to date has been on
desktop computers. An important next step in
this effort is implementation on a
high-performance computer for rapid analysis.
This work involves conversion of the existing
research-class, nonlinear statistical FORTRAN
code to a parallelized form that runs on a
high-performance cluster computer. The brain wave
data will be analyzed for seizure forewarning for
various values of the statistical parameters. The
goal is maximal total true rate, which is the sum
of true positives (correct prediction of a
seizure event) and true negatives (no forewarning
when no event occurs).
Objectives 1. Convert an existing
research-class, nonlinear statistical
FORTRAN code to a parallelized form that runs on
a high-performance cluster computer 2. Analyze
brain wave data for seizure forewarning for
various values of the statistical parameters 3.
Find a parameter set that maximizes the total
true rate
Seizure Prediction Process
2
1. EEG data gathered and sent to SeizAlert device
2(a). Artifact removal done and discrete points
generated
1
3
2(b). Analysis of dissimilarity
Methodology The program paradigm MPI (Message
Passing Interface) was used to convert the
existing code into a more parallelized form,
which was ran on a high-performance cluster.
3. Forewarning result
Fig.3. SeizAlert seizure prediction process
Results
How was it done?
(example)
Lets say,
T CPU time (seconds/program)
T-graph
FORTRAN MPI include mpif.h integer
myrank,a call MPI_INIT () call
MPI_COMM_RANK() call MPI_FINALIZE end
FORTRAN
Program Example integer a a 0 do I
1,10 a a1 end do print, a
Using MPI and running n processors,
, still (ideally)
But 10 times the output, when n 10
The CPU time was predicted to remain somewhat
constant, given the assumption that each
processor had identical calculating ability.
Also, examination of the forewarning summaries
made it possible to narrow the parameter space by
graphing the total true rate versus each
parameter.  By doing so a visible parameter
value region is shown, making it possible to see
which values generate higher total true rates.
Fig.1. Example of parallelized FORTRAN code using
MPI
As you can see, MPI isnt a program language in
itself, but is merely a software vehicle that
carries a code to and from processors and allows
numerous copies of a code to be executed
simultaneously.
Sent to processors
Gather results
Future Work
Initialization
Output
Biomedical examples include forewarning of
ventricular fibrillations (total true rate fT
5/5) and fainting (fT 2/2) from surface heart
waves (electro-cardiogram, ECG) detection of
sepsis onset (fT 23/23) from ECG, and breathing
difficulty from surface chest sounds (fT 2/2).
Industrial applications provide forewarning of
machine failures from motor power to tri-axial
acceleration.
Each job is calculated simultaneously
Fig.2. Parallel code using multiple processors
The Research Alliance in Math and Science program
is sponsored by the Mathematical, Information,
and Computational Sciences Division, Office of
Advanced Scientific Computing Research, U.S.
Department of Energy. The work was performed at
the Oak Ridge National Laboratory, which is
managed by UT-Battelle, LLC under Contract No.
De-AC05-00OR22725. This work has been authored by
a contractor of the U.S. Government, accordingly,
the U.S. Government retains a nonexclusive,
royalty-free license to publish or reproduce the
published form of this contribution, or allow
others to do so, for U.S. Government purposes.
OAK RIDGE NATIONAL LABORATORY U.S. DEPARTMENT OF
ENERGY
The author would like to thank Dr. Hively and Dr.
Nutaro for all the help on accomplishing my
project, as well as Dr. Deng of Alabama AM for
his encouragement and financial support to
participate in this program.