The scale of nowadays High Performance Computing (HPC) systems is the key element that determines the achievement of impressive performance, as well as the reason for their relatively limited reliability. Over the last decade, specific areas of the HPC research field have addressed the issue at different levels, by enriching the infrastructure, the platforms, or the algorithms with fault tolerance features. In this work, we focus on the rather-pervasive task of computing the solution of a dense, unstructured linear system and we propose an algorithm-based technique to obtain fault tolerance to multiple anywhere-located faults during the parallel computation. We particularly study the ways to boost the performance of the rollback-free recovery, and we provide an extensive evaluation of our technique w.r.t. to other state-of-the-art algorithm-based methods.

As large-scale linear equation systems are pervasive in many scientific fields, great efforts have been done over the last decade in realizing efficient techniques to solve such systems, possibly relying on High Performance Computing (HPC) infrastructures to boost the performance. In this framework, the ever-growing scale of supercomputers inevitably increases the frequency of faults, making it a crucial issue of HPC application development.A previous study [1] investigated the possibility to enhance the Inhibition Method (IMe) -a linear systems solver for dense unstructured matrices-with fault tolerance to single hard errors, i.e. failures causing one computing processor to stop.This article extends [1] by proposing an efficient technique to obtain fault tolerance to multiple hard errors, which may occur concurrently on different processors belonging to the same or different machines. An improved parallel implementation is also proposed, which is particularly suitable for HPC environments and moves towards the direction of a complete decentralization. The theoretical analysis suggests that the technique (which does not require check pointing, nor rollback) is able to provide fault tolerance to multiple faults at the price of a small overhead and a limited number of additional processors to store the checksums. Experimental results on a HPC architecture validate the theoretical study, showing promising performance improvements w.r.t. a popular fault-tolerant solving technique.

The ever-increasing size of High Performance Computing (HPC) systems inevitably causes an unwanted decrease of Mean Time Between Failures (MTBF). For this reason, over the last decade, much work has been done on the topic of fault tolerance for supercomputers. In particular, as large-scale linear algebra applications permeate many scientific fields, important efforts have been focused on the performance enhancements that could be provided by HPC infrastructures if reliable fault tolerant solutions are adopted. This article explores a popular topic of linear algebra (i.e., linear equation system resolution) and proposes an efficient, error-resilient approach based on an existing technique called Inhibition Method (IMe). Initially conceived to analyse complex electric circuits and later extended to solve linear systems, the original method is here enhanced with a simple, yet effective strategy to provide tolerance to single fail-stop recurring to neither checkpointing, nor rollbacks. Experimental results on a medium-scale HPC architecture show negligible overheads and promising performance improvements when compared with a popular fault-tolerant solving technique.