The work targets a stubborn problem: I/O variability. Modern HPC applications alternate between computation and I/O phases, and the I/O phases tend to be periodic, dominated by write operations such as checkpointing that occur in bursts synchronously across thousands of processes. Jobs that periodically write can account for a quarter to a third of the node-hours on large systems and over a third of all data written. But strictly periodic behavior is often valid only in specific time segments, and contention arises when different jobs compete for shared storage bandwidth. Existing I/O management approaches rely on predictive models built from past execution runs, which are error-prone precisely because I/O behavior varies between runs.

The research team, Ahmad Tarraf and Felix Wolf, enhanced their existing tool FTIO—Frequency Techniques for I/O—which predicts periodic I/O phases both online and offline by combining signal processing with outlier detection. FTIO samples the I/O bandwidth signal at a fixed sampling rate, applies the discrete Fourier transform to move it into the frequency domain, and uses the power spectrum to suppress noise from small-amplitude, high-frequency components. Outlier detection methods such as the Z-score, DBSCAN, or isolation forest then identify the dominant frequency. A Z-score beyond three typically flags an outlier, and the tool normalizes scores against an 80 percent tolerance to decide whether a signal is periodic with one candidate, weakly periodic with two, or not periodic at all.

The first enhancement gives users a richer description than a single dominant frequency. Because the Fourier decomposition expresses a signal as a sum of cosine waves, incorporating multiple frequency components makes the temporal characterization more accurate while keeping the model compact—only cosine waves are added. The team demonstrated this on the HACC-IO benchmark, which mimics the I/O phase of a cosmology code, executed with 384 processes on the Lichtenberg cluster at TU Darmstadt. Varying the number of extracted frequencies up to ten showed progressively more accurate representations of the writing phases.

The second refinement is Fourier fitting. Unlike the discrete Fourier transform, which returns the exact harmonic decomposition of a signal, Fourier fitting approximates the coefficients—amplitudes, phases, and frequencies—through non-linear least squares using SciPy’s curve_fit function, initialized with the top cosine waves from the DFT. This matters because frequencies lying between the discrete frequency bins of the DFT may fit the signal better. The researchers observed that Fourier fitting improves results when the signal contains noise from I/O variation, is not strongly periodic, or the sampling frequency is too low. On HACC-IO with 9,216 processes, selecting ten components already gave a good characterization, but Fourier fitting reduced the mean square error by 31.11 percent.

The centerpiece of the study is a hybrid approach that couples the discrete wavelet transform with the discrete Fourier transform. The wavelet transform sacrifices frequency resolution for time resolution, convolving the signal with scaled and translated versions of a mother wavelet—by default the Haar wavelet. Computed by successively passing the signal through high-pass and low-pass filters and downsampling by two at each decomposition level, the DWT produces detail and approximation coefficients that reveal which frequency ranges were active when. The team applies the DFT to the upsampled approximation coefficients, yielding a smoother signal with less variation than the raw trace.

From there, the method turns elegantly quantitative. If a dominant frequency is found, it determines the width of a sliding window and the number of samples inside it. The Pearson correlation between the dominant cosine wave from the DFT and the upsampled approximation coefficients is calculated inside the window, iterated across the length of the trace—sharing similarities with the continuous wavelet transform. Segments where the correlation stays positive are compared to the dominant cosine wave: segments shorter than their own period are discarded, and adjacent segments whose boundaries lie within one period are merged. The result is an array of start and end times when the signal can genuinely be characterized by the detected frequency.

The evaluation on the production Lichtenberg cluster, whose shared IBM Spectrum Scale file system achieves 150 GB/s, delivered concrete numbers. For the HACC-IO trace, FTIO found a dominant frequency of 0.055 Hz—corresponding to an 18.01-second period—but this representation was not valid for the entire 180-second run due to variations in the application. The hybrid analysis pinpointed that the frequency was valid inside the intervals 9.88 to 72.08 seconds and 106.38 to 168.28 seconds, with the entire process of computing the DWT, applying the sliding window correlation, and extracting results consuming just 4.2 seconds.

The team then demonstrated the approach on DLIO, an I/O benchmark for deep learning, configured with a modified ResNet50 setup that generates training data, trains the model, and performs 35.4 GB of checkpointing every epoch across ten epochs, using 96 MPI ranks on two nodes. With a sampling frequency of 10 Hz and a wavelet decomposition level of two, FTIO found in around four seconds that the dominant frequency of 0.0361 Hz—a 27.65-second period—was valid inside the interval 84.54 to 370.04 seconds. The researchers caution that further evaluation is needed to identify limitations, optimize the decomposition level, and enhance the correlation calculation at window boundaries.

The significance extends beyond one tool. Previous work showed that when FTIO’s predictions are combined with the I/O scheduler Set-10, system utilization rose by 26 percent and I/O slowdown fell by 56 percent. As exascale systems embrace modular supercomputing and increasingly dynamic, data-intensive workloads, knowing not just the period of an application’s I/O phases but exactly when those phases are valid could let schedulers orchestrate storage contention avoidance with unprecedented precision. The hybrid wavelet-Fourier characterization maintains a compact, interpretable behavioral model—unlike black-box machine learning predictors—while remaining efficient enough to run on production clusters. The datasets and reproduction instructions are publicly available, and the authors note the approach is promising, pending deeper examination of boundary calculations and parameter selections.

Subject of Research: HPC I/O performance prediction

Article Title: Improving I/O phase predictions in FTIO using hybrid wavelet-Fourier analysis

Article References: Tarraf, A., & Wolf, F. (2026). Improving I/O phase predictions in FTIO using hybrid wavelet-Fourier analysis. Frontiers in High Performance Computing, 3, Article 1638924. https://doi.org/10.3389/fhpcp.2025.1638924

Image Credits: AI Generated

DOI: 10.3389/fhpcp.2025.1638924

Keywords: high-performance computing, I/O phases, FTIO, wavelet transform, Fourier transform, signal processing, exascale computing, I/O scheduling, Improving, phase, predictions, hybrid