Write a Blog >>

The state-of-the-art yet decade-old I/O monitoring tool, Darshan, has been serving as one of the most fundamental subsystems to track and analyze the I/Os in high-performance computing (HPC) systems. One of the compelling challenges in Darshan that hinders its broader adoption lies in granularity: Darshan can only track the MPI and POSIX I/Os at the application level and brings little insight into the I/O activities inside the application: tasks, subroutines, etc. Besides, Darshan does not support those user-defined I/O libraries such as HDF5 and NetCDF.

We present Reflector, an I/O characterization tool, to support fine-grained I/O monitoring and report the I/O statistics at both the high-level and low-level I/O invocations. With Reflector, application users and system administrators can obtain deeper insights into their applications’ I/O activities not only implemented with POSIX- and MPI-I/O but also user-defined interfaces such as HDF5 and NetCDF. We have evaluated Reflector on both an on-premises 500-core HPC cluster and a leadership-class supercomputer at a U.S. DoE national lab. Preliminary results are promising as the system prototype incurs negligible performance overhead and clearly illustrates the I/O patterns and bottlenecks on multiple applications.