Mitigating I/O Bottlenecks for HPC Workloads

Modern HPC applications differ from traditional modeling and simulation workloads by using scale-out tasks like AI, big data analytics, deep learning, and complex workflows that have varying I/O access patterns. Hardware and software innovations are moving beyond the traditional two-tier model, introducing fast, composable storage tiers closer to computing resources. These increase the complexity of managing the conflicting needs of write- and read-intensive processes. This talk will focus on bringing out I/O bottlenecks when running emerging HPC workloads on modern HPC clusters and discuss potential solutions to address such I/O challenges.