Recently, the dramatic increase of the data amounts drives up the demand for data compression among HPC applications. Although many file systems and I/O middlewares have incorporated compression features, few high-level parallel I/O libraries support data compression due to the challenges of achieving scalable performance on HPC systems. This paper presents the design and implementation of the variable compression feature in the Parallel NetCDF library. Our design employs the same concept of chunking used by the HDF5 library, but we focus on enabling I/O aggregation across multiple requests to address the challenges on performance and scalability. We evaluate our solution using the I/O kernel of real-world scientific applications and analyze the impacts of data compression on parallel I/O performance. Our result suggests that handling multiple requests at once can significantly improve the parallel I/O performance on chunked and compressed data.

Accessing non-contiguous blocks in multiple array variables is a challenging I/O pattern for parallel applications to obtain good I/O performance. High-level I/O libraries such as HDF5 allow users to implement this pattern conveniently, but users have observed significant performance bottlenecks in the two-phase I/O implementation of MPI-IO. Recent studies have advanced the two-phase I/O performance by novel communication algorithms, but such improvements still have limitations. Two-phase I/O has to faithfully process inputs from high-level I/O libraries, so that implementation overheads can accumulate for improper usage of high-level I/O libraries. In this paper, we propose approaches for efficient usage of high-level I/O libraries that can circumvent major collective I/O overheads. We adopt a multi-dataset implementation of HDF5 dataset I/O to aggregate non-contiguous requests for array blocks and provide corresponding parameter assignment strategies. These approaches reduce the overheads caused by communication straggler effects in two-phase I/O. We show that our proposed methods can improve the parallel I/O performance up to 8× on two supercomputing systems for the HDF5 implementations of an I/O kernel extracted from climate simulation code compared with its baseline implementations.

Synchronous Stochastic Gradient Descent (SGD) with data parallelism, the most popular parallel training strategy for deep learning, suffers from expensive gradient communications. Local SGD with periodic model averaging is a promising alternative to synchronous SGD. The algorithm allows each worker to locally update its own model, and periodically averages the model parameters across all the workers. While this algorithm enjoys less frequent communications, the convergence rate is strongly affected by the number of workers. In order to scale up the local SGD training without losing accuracy, the number of workers should be sufficiently small so that the model converges reasonably fast. In this paper, we discuss how to exploit the degree of parallelism in local SGD while maintaining model accuracy. Our training strategy employs multiple groups of processes and each group trains a local model based on data parallelism. The local models are periodically averaged across all the groups. Based on this hierarchical parallelism, we design a model averaging algorithm that has a cheaper communication cost than allreduce-based approach. We also propose a practical metric for finding the maximum number of workers that does not cause a significant accuracy loss. Our experimental results demonstrate that our proposed training strategy provides a significantly improved scalability while achieving a comparable model accuracy to synchronous SGD.

The demands of increasingly large scientific application workflows lead to the need for more powerful supercomputers. As the scale of supercomputing systems have grown, the prediction of fault tolerance has become an increasingly critical area of study, since the prediction of system failures can improve performance by saving checkpoints in advance. We propose a real-time failure detection algorithm that adopts an event-based prediction model. The prediction model is a convolutional neural network that utilizes both traditional event attributes and additional spatio-temporal features. We present a case study using our proposed method with six years of reliability, availability, and serviceability event logs recorded by Mira, a Blue Gene/Q supercomputer at Argonne National Laboratory. In the case study, we have shown that our failure prediction model is not limited to predict the occurrence of failures in general. It is capable of accurately detecting specific types of critical failures such as coolant and power problems within reasonable lead time ranges. Our case study shows that the proposed method can achieve a F1score of 0.56 for general failures, 0.97 for coolant failures, and 0.86 for power failures.

Training Convolutional Neural Network (CNN) is a computationally intensive task, requiring efficient parallelization to shorten the execution time. Considering the ever-increasing size of available training data, the parallelization of CNN training becomes more important. Data-parallelism, a popular parallelization strategy that distributes the input data among compute processes, requires the mini-batch size to be sufficiently large to achieve a high degree of parallelism. However, training with large batch size is known to produce a low convergence accuracy. In image restoration problems, for example, the batch size is typically tuned to a small value between 16 ~ 64, making it challenging to scale up the training. In this paper, we propose a parallel CNN training strategy that gradually increases the mini-batch size and learning rate at run-time. While improving the scalability, this strategy also maintains the accuracy close to that of the training with a fixed small batch size. We evaluate the performance of the proposed parallel CNN training algorithm with image regression and classification applications using various models and datasets.

Geostatistical interpolation is the process that uses existing data and statistical models as inputs to predict data in unobserved spatio-temporal contexts as output. Kriging is a well-known geostatistical interpolation method that minimizes mean square error of prediction. The result interpolated by Kriging is accurate when consistency of statistical properties in data is assumed. However, without this assumption, Kriging interpolation has poor accuracy. To address this problem, this paper presents a new filtering-based clustering algorithm that partitions data into clusters such that the interpolation error within each cluster is significantly reduced, which in turn improves the overall accuracy. Comparisons to traditional Kriging are made with two real-world datasets using two error criteria: normalized mean square error(NMSE) and χ2 test statistics for normalized deviation measurement. Our method has reduced NMSE by more than 50% for both datasets over traditional Kriging. Moreover, χ2 tests have also shown significant improvements of our approach over traditional Kriging.