Using Darshan and CODES to Evaluate Application I/O Performance

Authors: Harsh Khetawat (North Carolina State University), Christopher Zimmer (Oak Ridge National Laboratory), Frank Mueller (North Carolina State University), Sudharshan Vazhkudai (Oak Ridge National Laboratory), Scott Atchley (Oak Ridge National Laboratory)

Abstract: Burst buffers have become increasingly popular in HPC systems, allowing bursty I/O traffic to be serviced faster without slowing down application execution. The ubiquity of burst buffers creates opportunities for studying their ideal placement in the HPC topology. Furthermore, the topology of the network interconnect can also affect the performance of the storage hierarchy for different burst buffer placement schemes. To that end, we create a reproducible framework that allows individual centers to develop their own models and evaluate performance based on their workload characteristics. We use CODES to create models that simulate the network and storage layers of an HPC system and Darshan traces for I/O replay. We augment the Darshan traces with synchronization primitives, and allow multi-dimensional scaling of traces to represent future workload characteristics. Finally, we evaluate the effect of network topology, storage architecture, and application I/O patterns on overall I/O performance.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF


Back to Poster Archive Listing