Performance Studies

We provide performance studies to support application users, developers, and system planners to assess an applications performance under differing architectures, software environments, and run environments

Service Overview

We know that hardware and software environments differ drastically in HPC centers and local clusters. The ability to assess the build process across various compilers, accelerators, and related drivers is a complex undertaking. The ability to optimize resource use on a given node and across a network and file system adds a second layer of complexity with interaction of the system resource scheduler adding a third.

Utilizing our SurveyPerf model, we can collaborate with you to identify, scope, and test your application(s) under scenarios that you identify to achieve your performance goals.

Study approach

We bring a team approach to a study project. Our team's expertise over the last four decades ranges from software systems to interaction with hardware components. We have expertise in Spack build systems, library integration, metric analysis, and much more. We provide a holistic approach where discussion and assessment is done as an extension of your team.

  • We start with understanding the issue that the user is facing, identifying the tasks and resources needed, and scoping a study plan. This initial step will provide the basis for the study contract and identify the tasks, resources, and timeframe.

  • The type of studies that we undertake range from a short term (weeks) concerted effort to provide build and run options to long term study contracts where we serve as a component to a deployment or integration team and provide performance assessment, comparative studies, or system impact assessments.

  • Another service we provide is integration of our Survey tool into Continuous Integration (CI) or testing frameworks. Our clients have found that integration of our Survey tool into the development or testing process provides insight of performance impacts that may normally not be noticed if tests results are binary (pass/fail).