Personalized Benchmarks for HPC Applications

Computational Science

Participants: Rob Ross, Phillip Carns

As applications became more data-intensive and the CPU-I/O gap grew, efficient HPC I/O stacks became critical for application performance. Benchmarks are particularly helpful for understanding, evaluating, and tuning all levels of the HPC I/O stack. While benchmarks’ use generally results in improved performance for the applications that inspired them, the benefits do not necessarily extend to the long tail of a platform’s workload. Because benchmarks are so labor-intensive to produce, they are only created for important applications and synthetic workloads focusing on one or two key aspects of I/O. Thus these benchmarks are of limited help to storage system designers and administrators who seek to understand the I/O behavior of their current or anticipated workloads, and to developers whose applications do not resemble existing benchmarks.

Our goal is to bridge this gap by devising methods to provide user-friendly personalized I/O benchmarks and analytics for the entire range of supercomputing jobs, applications, workloads, and stakeholders. We will design, test, and share an I/O boot camp curriculum, designed to help HPC users understand the range of I/O approaches available to them and incorporate instruction on using our analytics.

Personalized Benchmarks for HPC Applications