DescriptionHPC applications are generating increasingly large volumes of data (up to hundreds of TBs), which need to be stored in parallel to be scalable. Parallel I/O is a significant bottleneck in HPC applications, and is especially challenging in Adaptive Mesh Refinement (AMR) applications because the structure of output files changes dynamically during runtime. Data-intensive AMR applications run on the Cori supercomputer show variable and often poor I/O performance, but diagnosing the root cause remains challenging. Here we analyze logs from multiple levels of Cori's parallel I/O subsystems, and find bottlenecks during file metadata operations and during the writing of file contents that reduced I/O bandwidth by up to 40x. Such bottlenecks seemed to be system-dependent and not the application's fault. Increasing the granularity of file-system performance data will help provide conclusive causal relationships between file-system servers and metadata bottlenecks.