SCTuner: An Autotuner Addressing Dynamic I/O Needs on Supercomputer I/O Subsystems
Published in PDSW, 2021
In high-performance computing (HPC), scientificapplications often manage a massive amount of data usingI/O libraries. These libraries provide convenient data modelabstractions, help ensure data portability, and, most important,empower end users to improve I/O performance by tuningconfigurations across multiple layers of the HPC I/O stack. Wepropose SCTuner, an autotuner integrated within the I/O libraryitself to dynamically tune both the I/O library and the underlyingI/O stack at application runtime. To this end, we introduce astatistical benchmarking method to profile the behaviors of indi-vidual supercomputer I/O subsystems with varied configurationsacross I/O layers. We use the benchmarking results as the built-inknowledge in SCTuner, implement an I/O pattern extractor, andplan to implement an online performance tuner as the SCTunerruntime. We conducted a benchmarking analysis on the Summitsupercomputer and its GPFS file system Alpine. The preliminaryresults show that our method can effectively extract the consistentI/O behaviors of the target system under production load,building the base for I/O autotuning at application runtime.