Presenter: Kirsten Dawes
Co-Presenter: Elliott Ewing
Mentor: Hank Childs, Computer Science
Oral Presentation
Majors: Mathematics and Computer Science
As new computer architectures are developed, various concerns are created. Understanding scalability of performance is an important concern for many legacy programs on Intel’s Xeon Phi, a many integrated core architecture. These legacy programs use a significant amount of computational resources, which poses the issues of time, power, and memory consumption as the problem sizes increase. We will be looking at the scalability for the execution time of a program as the number of processors increase. Given this constraint, we should expect to see a decrease in the time of execution of a program using more processors. Exploring data-intensive workloads will allow us to be able to track how certain legacy programs mitigate the issue of scalability on this new architecture. We explored this new architecture by creating two competing mini applications, where each mini application replicated a variation on a legacy program called “VisIt”, which works on data intensive workloads. The two mini applications explored thread affinity and work affinity by implementing image processing filters with a data flow pipeline. We conducted experiments creating different size output and input images; number of image processing filters; type of image processing filters; and number of threads running. The experiments produced timing data based on CPU time as well as time per thread. With the quantitative data, we produced, we were able to conclude that data-intensive workloads like “VisIt” should produce excellent scalability on this new architecture.