Geoscience Reference
In-Depth Information
performance may include the following: that the full capacity of the GPU is not being utilised; that
the CPU-based measures are running into bottlenecks related to disc access in comparison to the
GPU method which runs in the memory of the graphics card; or that this represents the difference
between streaming and threading methods of allocating jobs to the processors.
The results from each of the algorithms are similar because GPU
k
-means uses the same stan-
dard
k
-means algorithm to cluster the data, albeit by running each instance of the
k
-means on a
single GPU core. The value of using such a parallel approach means that both exhaustive investiga-
tion of the number of clusters and different initial starting seeds can be undertaken.
3.9 CONCLUSIONS
Parallel computing has been applied in different fields to enhance the performance of algorithms
and computationally intensive tasks. Early objectives within the field of geography and GIS were
to make GIS run faster, but research has now moved beyond that goal and has also led to the
development of a few toolkits and implemented a few algorithms that exploit parallel computing
capabilities. Although there is evidence that the use of parallel computing in the field of geography
is increasing, use is still not very widespread. As discussed by Guan and Clarke (2010), it is not
clear if GIScience or geography more generally is really in an era of parallel computing or not as
most models are still based on sequential computing. The reasons for poor uptake can (partly) be
attributed to poor availability of parallel computing resources and lack of the necessary toolkits
and algorithms to build parallel applications (Clematis et al. 2003). It could also be attributed to
limited perceived need, limited interest in computing power and no real geographical champion.
Further, most reported applications offer little more than a combination of
exiting algorithms
and
data parallelism
, and finally, simultaneously thinking
in parallel
and
outside of the box
is a hugely
difficult activity. If learning curves are steep or if there are other barriers to accessibility, then
geographers may simply opt to settle for resolving small problems and/or suffer longer runtimes.
Lack of resources is becoming less of an issue with readily available GPU, although the deploy-
ment of this type of parallel computing to solve geographical problems is still rare. While GPU
began as a technology in which computer graphics cards were principally used to render visual
imagery, the development of GPGPU has enabled this technology to solve a variety of scientific
problems that are core to GC analysis. GPGPU technology is very scalable and very practicable
given the many multiple cores now available on high-specification graphics cards. Moreover, their
simple integration into an array structure allows the deployment of GPU alongside standard serv-
ers. These can be accessed remotely, making them highly suitable for implementation in data cen-
tres where specific analytical tasks require high-speed computation, as, for example, is required
for real-time financial analysis.
GPGPU is destined to become a growth area in GC, as well as more widely in both desktop and
web applications. The effects of wide adoption are likely to range from the speeding up of mundane
tasks associated with an operating system through to implementation of computationally intensive
algorithms to search for patterns and meaning hidden within massive data structures.
Clustering represents one important aspect of a chain of necessary spatial data operations which
will facilitate future web-based geodemographic classifications. The number and range of such
applications are likely to multiply in line with the greater availability of open spatial data pertaining
to populations and their interactions across space and the assembly of more real-time information
for use in decision-making. The illustrative application that we have developed here demonstrates
that GPGPU usage can be honed still further with the aid of standardisation procedures, such as
PCA. Our parallel
k
-means implementation using GPU shows significantly improved computation
efficiency when compared with the standard
k
-means clustering algorithm, and indeed, this scales
through addition of more GPUs, thus enabling further performance gains with appropriate hard-
ware. As such, we argue that the computation of an online geodemographic classification can be
achieved by combining PCA as the standardisation technique and a GPU-based parallel
k
-means
