210
A. Kubota et al.
Fig. 4 An example of
reconstructed
three-dimensional atom
images
3 Parallelization
As pointed out in Sect. 2, it is an issue that it takes long time to analyze the
crystal structure because reconstruction of two-dimensional atomic images and
observation of the images are repeated several times. We therefore try to improve
the reconstruction of three-dimensional atomic images by parallelizing DFT with
parallel programming language XcalableMP and parallel API OpenMP on multinode PC clusters.
OpenMP [7] is a parallel programming API which enables parallelization by
inserting directives in sequential programs. It enables parallelization of loops with
data parallelism easily and high performance on shared-memory machines can be
achieved. Similar to OpenMP, XcalableMP [10] is a parallel programming language
which also enables parallelization by inserting directives in sequential programs.
Data distribution among distributed-memory computing environment such as multinode PC clusters and supercomputers can be specified by inserting directives. Both
OpenMP and XcalableMP are designed so that directives related to parallelization
are ignored when the program is compiled to a sequential executable. Thus both
sequential and parallel programs are maintained in one common source code. In
addition, XcalableMP and OpenMP can be used at the same time for hybrid interand intra- node parallelization.
In this section, hybrid parallelization of reconstruction in two steps as follows:
1. parallelization of reconstruction of two-dimensional atomic images by OpenMP
on single PC node and
2. parallelization of reconstruction of three-dimensional atomic images by XcalableMP and OpenMP on multi-nodes
A. Kubota et al.
Fig. 4 An example of
reconstructed
three-dimensional atom
images
3 Parallelization
As pointed out in Sect. 2, it is an issue that it takes long time to analyze the
crystal structure because reconstruction of two-dimensional atomic images and
observation of the images are repeated several times. We therefore try to improve
the reconstruction of three-dimensional atomic images by parallelizing DFT with
parallel programming language XcalableMP and parallel API OpenMP on multinode PC clusters.
OpenMP [7] is a parallel programming API which enables parallelization by
inserting directives in sequential programs. It enables parallelization of loops with
data parallelism easily and high performance on shared-memory machines can be
achieved. Similar to OpenMP, XcalableMP [10] is a parallel programming language
which also enables parallelization by inserting directives in sequential programs.
Data distribution among distributed-memory computing environment such as multinode PC clusters and supercomputers can be specified by inserting directives. Both
OpenMP and XcalableMP are designed so that directives related to parallelization
are ignored when the program is compiled to a sequential executable. Thus both
sequential and parallel programs are maintained in one common source code. In
addition, XcalableMP and OpenMP can be used at the same time for hybrid interand intra- node parallelization.
In this section, hybrid parallelization of reconstruction in two steps as follows:
1. parallelization of reconstruction of two-dimensional atomic images by OpenMP
on single PC node and
2. parallelization of reconstruction of three-dimensional atomic images by XcalableMP and OpenMP on multi-nodes