Algorithmic approaches to low overhead fault detection for sparse linear algebra

Abstract

The increasing size and complexity of HighPerformance Computing systems is making it increasingly likely that individual circuits will produce erroneous results, especially when operated in a low energy mode. Previous techniques for Algorithm -Based Fault Tolerance (ABFT) [20] have been proposed for detecting errors in dense linear operations, but have high overhead in the context of sparse problems. In this paper, we propose a set of algorithmic techniques that minimize the overhead of fault detection for sparse problems. The techniques are based on two insights. First, many sparse problems are well structured (e.g. diagonal, banded diagonal, block diagonal), which allows for sampling techniques to produce good approximations of the checks used for fault detection. These approximate checks may be acceptable for many sparse linear algebra applications. Second, many linear applications have enough reuse that preconditioning techniques can be used to make these applications more amenable to low-cost algorithmic checks. The proposed techniques are shown to yield up to 2x reductions in performance overhead over traditional ABFT checks for a spectrum of sparse problems. A case study using common linear solvers further illustrates the benefits of the proposed algorithmic techniques.Index Terms-ABFT, sparse linear algebra, numerical methods, error detection

Mechanisms and Evaluation of Cross-Layer Fault-Tolerance for Supercomputing

“…By decoupling, we assume a more relaxed model that allows incorrect results to be written. With algorithmic detection techniques [51], we have no natural instruction-precise "immediate" retry so we must identify a well defined point. Also, this decoupling allows our results to consider techniques like Truffle [16], which steers instructions into low-energy approximate or high-energy precise pipelines.…”

A framework for evaluating comprehensive fault resilience mechanisms in numerical programs

“…Probabilistic modeling has been used by Chung et al [7] to help compute the expected recovery time, which cannot be measured easily for very large scale programs. Sloan et al [30] have discussed the use of algorithmic checks over sparse linear algebra kernels and focused mainly on reducing false positive and false negative in error detection.…”

Characterization of Impact of Transient Faults and Detection of Data Corruption Errors in Large-Scale N-Body Programs Using Graphics Processing Units

“…Algorithm-based fault tolerance (ABFT) [19] is a superior example of customized protection, because it offers high detection coverage and low runtime overhead in fundamental linear algebra operations and other matrix operations (e.g., LU decomposition [20], sparse matrix multiplication [21], and iterative liner solvers [22]). With ABFT, rewriting of an application program usually takes a large amount of manual effort, while our approach automatically customizes error detectors by itself using an unsupervised learning procedure once the protection target states (which are usually GPU kernel output data) have been identified by users.…”