Multi-threading ASKAPSoft Synthesis Imaging

CI: – Dr. Daniel Mitchel


Aim: to maximise the scientific output of the real-time ASKAP survey pipelines by adding multithreading to two algorithms that dictate the performance of ASKAPSoft, a modular synthesis imaging package designed primarily to serve as the underlying pipeline tasks. A further aim is to maximise the scientific output of ASKAPSoft in general, as it is also used for ASKAP early science processing and increasingly as a general-purpose package.

Significance: the project will increase the productivity of the ASKAP surveys and the flexibility to port the software to new hardware. ASKAPSoft algorithms are well suited to its underlying MPIbased data parallelism. However, the worker processes can require 10s of GB of memory, making it impractical to use all 20 cores per node of the dedicated Cray XC30 cluster, Galaxy, at the Pawsey Supercomputing Centre. A combination of MPI and OpenMP offers a better match to the system. Furthermore, the workers alternate processing with a master process, forming a bottleneck in an otherwise highly distributed system. The master requires more memory than the workers and in turn a full node to itself; it is an ideal candidate for multi-threading. Finally, Galaxy is set to be replaced in the next year or two, and this project will provide greater flexibility during procurement and better optimisation on the new system.

Expected outcomes: ASKAPSoft will have an additional level of parallelism that allows it to be tuned to a given cluster, and identified optimal settings for Galaxy will be worked into the standard ASKAP pipelines.

Categories