In-Line vs. In-Transit In Situ: Which Technique to Use at Scale?

Abstract

In situ visualization is increasingly necessary to address I/O limitations on supercomputers. With the increasing heterogeneity of supercomputer design, efficient and cost effective use of resources is extremely difficult for in situ visualization routines. In this work, we present a time and cost analysis of two different classes of common visualization algorithms in order to determine which in situ paradigm (in-line or in-transit) to use at scale, and under what circumstances. We explore a high computation and low communication algorithm, as well as a low computation and medium communication algorithm. We use 255 individual experimental runs to compare these algorithms performance at scale (up to 32,768 cores in-line and 16,384 core in-transit) with a running simulation. Finally, we show that — contrary to community belief — in-transit in situ has the potential to be both faster and more cost efficient than in-line in situ. We term this discovery Visualization Cost Efficiency Factor (VCEF), which is a measure of how much more performant in-transit in situ is on a smaller subset of nodes than in-line in situ is at the full scale of a simulation. Our results for these algorithms showed in-transit VCEF values of up to 8X at our highest concurrencies.

This dissertation includes previously published co-authored material