Description: This seminar will present an overview of the Empirical Roofline Toolkit (ERT) and discuss the input configuration required to generate the plots.  The aim is to provide users with a tool to relate performance to a system’s limitations in a quantitative sense.  Roofline plots are useful in answering the ultimate question in performance optimization, “Is the application truly getting good performance?”

This seminar is a follow-on to Introduction to Roofline Plots.  It is suggested that attendees unfamiliar with roofline modeling review that training seminar.

The ERT is a software tool to analyze and visualize the measured performance characteristics of computer systems, particularly focusing on memory bandwidth and computational capabilities.  The resulting output is an ERT graph that is based on the Roofline performance model, which provides an intuitive way to understand the relationship between computational performance, memory bandwidth, and arithmetic intensity.

CUI

Description: This seminar will introduce tools for profiling Python scripts to analyze performance and monitor memory usage. Demonstrations on an HPC system will cover how to profile basic, multi-threaded, and MPI scripts. We will also look at various ways to visualize profile data and a few tools for profiling in Jupyter Notebooks.

CUI

Description: This seminar will present an overview of the Roofline Model. The roofline model is model used to analyze the performance of computing architectures. It provides a way to assess the potential performance of an application by comparing its arithmetic intensity (the ratio of floating-point operations to memory operations) against the hardware's capabilities.

Roofline plots are useful in answering the ultimate question in performance optimization, “Is the application truly getting good performance?” By using the roofline model, developers can better understand how their code performs relative to hardware characteristics. Performance boundaries are depicted using ‘roofs’ and ‘lines’:

(1) The "roof" indicates the maximum achievable performance based on the system's peak computational capabilities; and

(2) The "lines" below the roof represent the performance limited by memory bandwidth.

Performance measurements are characterized as compute bound or memory bound. Identifying bottlenecks using these characterizations are useful in guiding optimization efforts.

CUI

Description: This seminar will discuss the Valgrind tool suite that provides a variety of tools for debugging and profiling serial and parallel applications. This includes the popular Memcheck tool which helps users find and fix difficult to locate memory errors and identify issues that may produce errors at times. Performance analysis tools including the Cachegrind and Callgrind tracing tools and the Massif heap profiling tool will also be discussed. Examples will be provided showing how to use these tools and common issues these tools can help resolve will be discussed.

CUI

Description: Linux tools can provide a wealth of information about: an HPC node’s hardware; CPU and memory use over time; and resource use by individual processes. This information helps users tune their applications. This course covers HPC CPUs, available tools, and several methods of adapting application deployment.

CUI