Highlight

Comparing the Efficiency of In Situ Visualization Paradigms at Scale

Scaling performance graph of in-line and in-transit use cases.
Total cost in node-seconds to run each simulation configuration using no visualization, in-line visualization, and three configurations of in-transit visualization (with 12%, 25%, and 50% additional resources respectively). Demonstrating that in-transit resources can pay for themselves in certain configurations. (Image Credit James Kress)

Achievement

We demonstrated that in-transit techniques can be faster AND cheaper at scale for certain visualization configurations. We did this by studying the performance and cost characteristics of the two major in situ techniques, in-line and in-transit.

Significance and Impact

It has traditionally been thought that in-transit visualization, while potentially providing a time savings to a simulation run, is more expensive to use than in-line techniques due to the additional resources required. We show that this is not always the case. In fact, it is possible to have in-transit techniques be both faster and more cost efficient vs. in-line methods at large scale.

 

Research Details

  • Utilized a proxy application running on Titan using VTK-m for visualization and ADIOS for in-transit data movement
  • Characterized the total cost of both in-line and in-transit, creating a cost model that shows when one outperforms the other at a given scale.
  • Exploring an extension to the work that uses more algorithms and larger scales.

Overview

With this work we looked at the overheads of in-transit and in-line visualization at various scales. We demonstrate that in-transit techniques can provide not only a time savings when doing visualization at large scale, but that by dedicating resources exclusively for visualization, it is possible to have the overall resource cost of using visualization be lower vs. in-line techniques. This is a significant finding, as it has generally been regarded in the community that in-transit techniques while having the potential to have a faster time to results, cost more to use than in-line methods. Our study shows that as inter process communication increases at scale, it is possible to dedicate very large resources to in-transit visualization and still have the overall run cost less than comparable in-line methods.

 

Comparing the Efficiency of In Situ Visualization Paradigms at Scale 
James Kress, Matthew Larsen, Jong Choi, Mark Kim, Matthew Wolf, Norbert Podhorszki, Scott Klasky, Hank Childs, and David Pugmire 
ISC High Performance Conference, Frankfurt, Germany, June 2019 

https://link.springer.com/chapter/10.1007/978-3-030-20656-7_6