The International Conference for High Performance Computing, Networking, Storage and Analysis
Comparing Decoupled I/O Kernels versus Real Traces in the I/O Analysis of the HACC Scientific Applications on Large-Scale Systems.
Student: Sean McDaniel (University of Delaware)
Supervisor: Hai Ah Nam (Oak Ridge National Laboratory)
Abstract: Synthetic benchmarks are often inadequate in comparison to real-world applications as large-scale end-to-end system stressors. Current solutions fail to put systems under significant stress and running real-world applications for testing and tuning is not always feasible because requiring increased resources and time. Lack in adequate solutions for stressing systems has increased efforts in decoupling I/O from applications to be used as benchmarks. Currently there is not a general methodology for extracting I/O and communication codes from scientific applications. In this work we compare the real I/O traces of a cosmology framework the Hardware/Hybrid Accelerated Cosmology Codes (HACC) with the traces of its decoupled I/O kernel (HACC-I/O). We validate results of the decoupled I/O kernel versus the HACC's I/O of simulations to quantify whether the kernel reliably mimics the I/O patterns of the real application. We outline similarities and discrepancies; the latter open exciting research opportunities in the HACC's IO tuning.