International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
[! Compute EGM2008 geoid heights by spherical harmonic analysis |
| latitude count / node count
longitude count / CPU count
Han | [aux | |
gather results computed in whole nodes
| Fit EGM2008-derived geoid heights to GPS/leveling data using LSC vik
| correlation count/ node count (A)
e] et o o oboe
(A)/ CPU count
ea li ons han]
gather results computed in whole nodes
x
| Find optimum correlation length and output fit results |
Figure 2. Parallelization model flow chart of SHA and LSC
4, PARALLELIZATION MODELING AND
IMPLEMENTATION
4.1 Parallelization modelling
This research performed parallelization modeling by applying
hybrid approach to analyze accuracy of geoid heights that is
finally determined by fitting the EGM2008-derived geoid
heights to geometric geoid heights calculated from
GPS/levelling data and determine optimum correlation length.
The parallelization modeling consists of SHA and LSC (Figure
2). And a task of the parallelization model is separated into one
cluster master and cluster nodes. The cluster master performs to
allocate tasks and gather results in each node. And the cluster
nodes perform tasks of allocated regions.
First, the master node is to allocate EGM2008 coefficients,
working scopes to compute EGM2008-dervied geoid heights to
each cluster node. Here, generally, working scopes used for
distributed parallel processing are divided with grid shapes but
their associated Legendre function values are same in same
latitude (Xiao and lu, 2007). Therefore, the working scopes can
be divided based on latitude. Next, the master node performs to
gather EGM2008-dervied geoid heights and then distributes the
EGM2008-derived geoid heights and working scopes for LSC
tasks to each cluster node. Next, each cluster node carries out
LSC fitting using the EGM2008-derived geoid heights,
geometric geoid heights calculated from GPS/levelling data and
correlation length and explores correlation length which has
minimum standard deviation at corresponding node. Finally, the
cluster master performs to gather optimum correlation length in
cach node and then chooses final optimum correlation length.
And hybrid MPI and OpenMP application to carry out SHA and
LSC is ported so that it can be executed on Linux platform
written with Fortran 77 language (Forsberg et al., 2003; Rapp,
1982; Yecai, 1994). These are implemented by C++ using
MPICH2 and OpenMP library and also compiled by GNU C
Compiler.
4.0 Computing platform
Diskless-based PC cluster system is implemented for distributed
parallel computing using hybrid MPI and OpenMP approach
proposed in this paper (Table 1). This cluster system consists of
1 cluster master and 16 cluster nodes, and its network
environments is made of 100MB bps switch hub. Operating
system software for cluster master used Community ENTerprise
Operating System (CentOS) 6.2 x86 64bit, with GCC 4.4.6
compiler used for compilation of code and MPI library used
MPICH2 1.2.1 (http://www.mcs.anl.gov/mpi/mpich2). Cluster
nodes are structured to be driven by diskless method using
Perceus 1.6.1 (http://www.perceus.org/) which is a program
package made by Infiscale for driving diskless cluster while
each node of cluster system is structured in a way that OS image
stored at cluster master is connected by PXE network booting
method.
Table 1. Summary of the diskless-based PC cluster system
Cluster Master Cluster node
No. of nodes 1 16
Processor Intel® Core™ i3- Intel® Pentium® Dual
Model 2120 CPU E2200
# of Cores 2 2
# of Threads 4 2
Processor speed 3.3 GHz 2.2 GHz
Memory 4 GB 2GB
Ethernet card Realtek gigabit ethernet card
5. RESULTS AND DISCUSSION
Experiment of the parallelization modeling in this paper is
carried out in the diskless cluster system consisting of low-
performance computers by hybrid MPI and OpenMP
application. And this research carried out SHA of EGM2008 for
around the Korean peninsula using parallization modeling
applying hybrid approach developed in this research. SHA is
performed at latitude 32°N-43°N and longitude 123°E-132°E
by applying maximum degree and order (2,190) of EGM2008.
From the results of analysis, the EGM2008-derived geoid
heights are calculated for a total of 357,601 grid points with 1
arc minute of interval for latitude and longitude respectively.
And LSC was performed between 0.1 and 150 km by 0.1 km
interval. And broadcast time of EGM2008 and the EGM2008-
derived geoid heights runtime for SHA of EGM2008, runtime
to find optimum correlation length, and total computation time
are measured (Figure 3). 2 processors means 1 node and 1
processor means an environment without node i.e. test result by
serial algorithm at single computer not at cluster system.
214
tim
