Improvement in interactive remote in situ visualization using SIMD-aware function parser and asynchronous data I/O

Abstract

An in situ visualization system based on the particle-based volume rendering offers a highly scalable and flexible visual analytics environment based on multivariate volume rendering. Although it showed excellent computational performance on the conventional CPU platforms, accelerated computation on the latest many core platforms revealed performance bottlenecks related to a function parser and particles I/O. The function parsers handle multidimensional transfer functions, but conventional implementation was not optimized for wide SIMD widths. The I/O bottleneck comes from the latency of output of particle data files. In this paper, we develop a new SIMD-aware function parser and an asynchronous data I/O method based on task-based thread parallelization. The particle generation process is optimized by loop blocking to take advantage of the new function parser. Numerical experiments on the Oakforest-PACS, which consists of 8208 Intel Xeon Phi7250 (Knights Landing) processors, demonstrate an order of magnitude speedup with keeping improved strong scaling up to $$\sim 100\,\hbox {k}$$ cores.