Please join us for the CUP-ECS Colloquium, “Determining Proxy Application Fidelity using ML-Based Similarity Methods,” presented by Jeanine Cook of Sandia National Laboratories. This event will take place Friday, February 5, 2021 at 12 p.m. PST/1 p.m. MST/ 3 p.m. EST on Zoom.
Friday, February 5, 2021 3 p.m. EST
Determining Proxy Application Fidelity using ML-Based Similarity Methods
Jeanine Cook, Sandia National Laboratories
Abstract: Proxy applications are smaller models of a larger parent application that are targeted to represent particular characteristics such as programming model or node, memory, or communication behavior of the parent. They are widely used in hardware/software co-design, as code development tools, and in system acquisition as performance benchmarks. We often accept, qualitatively, that proxies are faithful representations of their parents, but this has little quantitative validation. In this work, we present a machine learning-based method called cosine similarity to quantitatively understand the correspondence of proxy to parent with respect to their underlying node and memory hardware behavior. We present a similarity analysis using a suite of applications across two different system architectures that shows where a proxy is similar to a parent and where it is not with respect to node behavior, and also present a redundancy analysis to show where proxies (and applications) are similar to each other. The results show that most proxies are good representations of their parents at least in some node/memory component behavior but there are outliers, which means care must be taken in choosing a proxy suite that truly covers the desired behaviors.