Abstract: Radio telescopes like the VLA/ngVLA/DSA2000/EHT use an indirect imaging technique to synthesis apertures much larger than the size of single antenna aperture. This enables angular resolution of sub-milliarcsecs and unprecedented sensitivity to cosmic signals enabling ground breaking scientific discoveries. Use of this imaging techniques however comes at the cost of applying computationally expensive algorithms to vast quantities of data to reconstruct an image of the sky. Estimate for the next-generation telescopes are in the many 10s of PFLOP/sec of computing required to process 100s of Tera Bytes of data. Challenges posed by the combination of high-computing and large-data needs makes the problem of indirect imaging significantly different from other HPC applications which require one or the other, but not both together.
In this talk, we will give a quick introduction to aperture synthesis imaging and motivate the combined challenge of high-computing and large-data in general. To mitigate the computational challenge, we have recently deployed the computing hot-spot in our processing on GPUs with the software built on the Kokkos framework. We will describe its use in our algorithms architecture, show the resulting runtime gains, and discuss the outstanding challenges and problems that still need solutions.
Abstract: Radio telescopes like the VLA/ngVLA/DSA2000/EHT use an indirect imaging technique to synthesis apertures much larger than the size of single antenna aperture. This enables angular resolution of sub-milliarcsecs and unprecedented sensitivity to cosmic signals enabling ground breaking scientific discoveries. Use of this imaging techniques however comes at the cost of applying computationally expensive algorithms to vast quantities of data to reconstruct an image of the sky. Estimate for the next-generation telescopes are in the many 10s of PFLOP/sec of computing required to process 100s of Tera Bytes of data. Challenges posed by the combination of high-computing and large-data needs makes the problem of indirect imaging significantly different from other HPC applications which require one or the other, but not both together.
In this talk, we will give a quick introduction to aperture synthesis imaging and motivate the combined challenge of high-computing and large-data in general. To mitigate the computational challenge, we have recently deployed the computing hot-spot in our processing on GPUs with the software built on the Kokkos framework. We will describe its use in our algorithms architecture, show the resulting runtime gains, and discuss the outstanding challenges and problems that still need solutions.