Getting Started

Documentation

Glish

Learn More

Programming

Contact Us

Version 1.9 Build 1488

News	FAQ
Search	Home

Subsections

• Priorities
• Targets

Parallelization and high-performance computing

Priorities

Scientific application

The parallelization effort needs to demonstrate a scientifically useful capability to address the most challenging problems in radio astronomy where supercomputer resources are required. These include wide-field imaging problems at low observing frequencies, mosaicing and the largest VLBI observations, amongst others. In addition, this also includes new algorithms which have not been widely used to date due to limited computing resources.

Parallelization infrastructure

A central goal of the parallelization effort is to ensure that infrastructure is developed within the AIPS++ system as a whole to support parallel and distributed computing without expensive ad hoc modifications. This requires that the parallelization infrastructure be compatible with the overall project design, and also that the mainstream project development consider parallelization when implementing algorithms. It is also imperative that the parallelization capabilities be presented using the same user interface as the conventional package.

High-performance computing in AIPS++

The parallelization effort has a strong vested interest in the serial performance of AIPS++ for problems of the largest size, which are defined to be those with exceptional I/O, memory or CPU requirements. It is considered the responsibility of this group to profile the serial performance in these specialized cases, make any changes required to support these large problem sizes, and optimize overall serial performance in these cases.

Targets

Cluster Linux and IRIX build maintenance

Continued maintenance of the existing NCSA/NRAO builds under IRIX and on the AHPCC Linux cluster. (WY, H, 3 wk, DM, H, 3 wk).

Key project processing

Processing of five key projects (including at least one each of mosaiced, wide-field or large spectral line). Candidates include the existing M33 dataset (Westpfahl), TXCam (Kemball), and a selection of low-frequency VLA projects in A-configuration. Generation of user liaison documentation and user support at NCSA. (RP, H, 2 wk, AS, H, 2 wk, WY, H, 2 wk).

Complete multi-field parallelization

Complete implementation of a prototype parallelization for mosaiced or wide-field imaging. Candidates include field- or facet-based gridding, model prediction or residual image computation. (KG, H, 4 wk).

Parallelization of other deconvolution methods

Extend the Clark CLEAN parallelization to other deconvolution algorithms. (KG, H, 4 wk).

Initial parallel I/O implementation

Implementation of multi-process I/O on the same file, multi-iterator access in a single process, and ROMIO asynchronous I/O using MPI-2. Evaluation of performance in these cases. (WY, H, 4 wk).

Complete NT migration document

Combine work by P. Cortes in NT migration to formalize incorporation of these changes in the main code distribution (WY, H, 1 wk, PC, H, 3 wk).

Modify test suite for large problem size

Modify the current bigimagertest to use simulated data. (DM, H, 1 wk).

Cluster parallelization

Expand evaluation of parallelization methods specific to cluster architectures, with a specific focus on Linux systems. Port AIPS++ to IA64 when available. (PC, H, 7 wk)