Characterizing and Predicting the I/O Performance of HPC Applications Using a Parameterized Synthetic Benchmark H. Shan, K. Antypas, J. Shalf CRD/NERSC, Lawrence Berkeley National Laboratory Berkeley, CA 94720 {hshan, kantypas, jshalf@lbl.gov} Abstract The unprecedented parallelism of new supercomputing platforms poses tremendous challenges to achieving scal- able performance for I/O intensive applications. Performance assessments using traditional I/O system and component benchmarks are difficult to relate back to application I/O requirements. However, the complexity of full applications motivates development of simpler synthetic I/O benchmarks as proxies to the full application. In this paper we exam- ine the I/O requirements of a range of HPC applications and describe how the LLNL IOR synthetic benchmark was chosen as suitable proxy for the diverse workload. We show a procedure for selecting IOR parameters to match the I/O patterns of the selected applications and show it can accurately predict the I/O performance of the full applica- tions. We conclude that IOR is an effective replacement for full-application I/O benchmarks and can bridge the gap of understanding that typically exists between stand-alone benchmarks and the full applications they intend to model. 1 Introduction The advent of petascale computing is leading to High-End Computing platforms of unprecedented concurrencies. This daunting level of parallelism will pose enormous challenges for future I/O systems that must support efficient and scalable data movement between disks and distributed memories. In order to guide the design of new I/O systems, it is necessary to gain a better understanding of applications’ I/O requirements. However, it can be impractical to run full applications for testing and evaluation of emerging I/O solutions, motivating the need for a compact synthetic benchmark capable of modeling the I/O access patterns of a diverse workload. Such a benchmark must not only be able to model application behavior, it must also be able to predict application performance to prove it is a suitable proxy for the full application codes it replaces. Synthetic I/O benchmarks tend to offer performance metrics that relate directly to system or hardware components of the disk subsystem (eg. random vs. contiguous reads or writes), but are extraordinarily difficult to relate back to application requirements. By contract, full application benchmarks enable head-to-head comparisons of the effective performance delivered to applications, but offer very little insight or diagnostic capability because they do not isolate any specific system or hardware component of the underlying parallel filesystem. This creates a gap in understanding between the hardware-oriented synthetic benchmarks and the information that can be gathered from full applications.