Skip to main content
eScholarship
Open Access Publications from the University of California

Toward Resilience and Data Reduction in Exascale Scientific Computing

  • Author(s): Liang, Xin
  • Advisor(s): Chen, Zizhong
  • Cappello, Franck
  • et al.
Abstract

Because of the ever-increasing execution scale, reliability and data management are becoming more and more important for scientific applications. On the one hand, exascale systems are anticipated to be more susceptible to soft errors ,e.g. silent data corruptions, due to the reduction in the size of transistors and the increase of the number of components. These errors will lead to corrupted results without warning, making the output of the computation untrustable. On the other hand, large volumes of highly variable data are produced by scientific computing with high velocity on exascale systems or advanced instruments, and the I/O time on storing these data is prohibitive due to the I/O bottleneck in parallel file systems. In this work, we leverage algorithm-based fault tolerance (ABFT) and error-bound lossy compression to tackle the two problems, in order to support efficient scientific computing on exascale systems.

We propose an efficient fault tolerant scheme to tolerant soft errors in Fast Fourier Transform (FFT), one of the most important computation kernels widely used in scientific computing. Traditional redundancy approaches will at least double the execution time or resources, limiting the usage in practice because of the large overhead. Previous works on offline ABFT algorithms for FFT mitigate this problem by providing resilient FFT with lower overhead, but these algorithms fail to make progress in vulnerable environments with high error rates because they can only detect and correct errors after the whole computation finishes. We propose an online ABFT scheme for large-scale FFT inspired by the divide-and-conquer nature of the FFT computation. We devise fault tolerant schemes for both computational and memory errors in FFT, with both serial and parallel optimizations. Experimental results demonstrate that the proposed approach provides more timely error detection and recovery as well as better fault coverage with less overhead, compared to the offline ABFT algorithm.

To alleviate the I/O bottleneck in the parallel file systems, we work on a prediction-based error-bounded lossy compressor to significantly reduce the size of scientific datasets while retaining the accuracy of the decompressed data, with adaptive prediction algorithms and compression models. We first propose a regression-based predictor for better prediction accuracy than traditional approaches under large error bounds, followed by an adaptive algorithm that dynamically selects between the traditional Lorenzo predictor and the proposed regression-based predictor, leading to very high compression ratios with little visual distortion. We further unify the prediction-based model and transform-baed model by using transform-based compressors as a predictor, with novel optimizations toward efficient coefficient encoding for both the two models. The proposed adaptive multi-algorithm design provides better compression ratios given the same distortion, significantly reducing storage requirements and I/O time.

We further adapt the compression algorithms and compressors to different requirements and/or objectives in realistic scenarios. We leverage a logarithmic transform to precondition the data, which turns a relative-error-bound compression problem into an absolute-error-bound compression problem. This transform aligns two different error requirements while improving the compression quality, efficiently reducing the workload for compressor design. We also correlate the compression algorithm with system information to achieve better I/O performance compared to traditional single compressor deployment. These studies further improve the efficiency of lossy compression from the perspective of efficient I/O in the context of scientific simulation, making scientific applications running on exascale systems more efficient.

Main Content
Current View