Our partner the Goethe University Frankfurt am Main had a paper accepted for 4th Workshop on on Extreme-Scale Storage and Analysis (ESSA’23), held in conjunction with IPDPS’23.
Since the performance of an application in modern HPC and supercomputing systems is not only limited by computations and memory accesses, new techniques are needed to characterize the performance of the system or application. In this preliminary work, we propose an empirical Roofline model and corresponding workflow for I/O workload analysis that can be adapted in the future for a multidimensional score, allowing to characterize application performance for different HPC systems. Regarding the I/O aspect, our empirical Roofline model focuses on the most commonly used performance metrics, I/O operations per second (IOPS) and I/O bandwidth. By applying the Roofline modeling, the I/O performance of an application can be intuitively characterized and possible performance bottlenecks can be identified without any deeper knowledge of the I/O stack. Providing a case study based on different application I/O kernels, we demonstrate that our empirical Roofline model is suitable to characterize the performance for a variety of applications.