Previous Reports
Research Reports

High-Performance Interconnects: “Performance” is Still HPC’s Middle Name

by Addison Snell, Laura Segervall
for Intersect360 Research
July 2015

In this study, Intersect360 Research conducted a custom, sponsored, broad-based survey of end users in the High Performance Computing (HPC) industry to assess the most important characteristics of system interconnects for HPC systems. These survey questions were answered by 197 unique respondents, spanning industry, government, and academia worldwide.
HPC end users were asked to answer two questions. In the first, they rated nine different features related to system interconnects on their importance, on a five-point scale. In the second, they rated potential barriers to achieving greater scalability, again on a five-point scale.
The results of the study were clear and consistent across all segments, that system bandwidth and interconnect latency are the two paramount system interconnect features for HPC, with the highest average scores and the highest percent of users rating them a 4 or a 5 on the scale.
Other interesting findings lurked below the surface. All of the features listed had average scores above the midpoint of the scale and at least 43% of users rating them a 4 or 5, meaning that any site might have additional features that are as important as bandwidth and latency. For public-sector (academia and government) sites, the most likely also-important feature was the implementation of programming model (e.g., MPI) semantics in hardware, whereas commercial end users favored forward and backward compatibility and standardization on open-source software.
The results of this study highlight the enduring, paramount importance of performance—the P in HPC. Whereas enterprise IT applications may be predominantly driven by reliability and standards, the nature of HPC applications is to strive to new pinnacles of discovery, insight, or achievement. Scientific and engineering problems can be run at higher fidelity, with increased realism, with more degrees of freedom in the model. You don’t need to build the same bridge twice; you go build a harder bridge. The questions in this study probe what it takes to pursue the goal of an ever-receding horizon. Without performance, nothing else matters very much.
The path to high performance may incorporate many specific features that were beyond the scope of the survey. Smarter interconnects, for example, based on intelligent routing or the ability to offload MPI semantics, can translate directly to greater application performance. These detailed features must be demonstrated as a one-step remove from a straightforward performance claim, but they are linked, and in the future we could see an increase in the direct demand for features like offloading and interconnect operating systems to optimize performance at scale.

For full access to our HPC research studies, analysis, custom research and consulting, contact us directly.

» 888-256-0124
» Subscribe to our mailing list

Sign in to view premium content
Remember me on this computer
Become a client
Get access to our full reports