Several recent publications have shown that hardware faults in the memory subsystem are commonplace. These faults are predicted to become more frequent in future systems that contain orders of magnitude more DRAM and SRAM than found in current memory subsystems. These memory subsystems will need to provide resilience techniques to tolerate these faults when deployed in high-performance computing systems and data centers containing tens of thousands of nodes. Therefore, it is critical to understand the efficacy of current hardware resilience techniques to determine whether they will be suitable for future systems. In this paper, we present a study of DRAM and SRAM faults and errors from the field. We use data from two leadership-class high-performance computer systems to analyze the reliability impact of hardware resilience schemes that are deployed in current systems. Our study has several key findings about the efficacy of many currently deployed reliability techniques such as DRAM ECC, DDR address/command parity, and SRAM ECC and parity. We also perform a methodological study, and find that counting errors instead of faults, a common practice among researchers and data center operators, can lead to incorrect conclusions about system reliability. Finally, we use our data to project the needs of future large-scale systems. We find that SRAM faults are unlikely to pose a significantly larger reliability threat in the future, while DRAM faults will be a major concern and stronger DRAM resilience schemes will be needed to maintain acceptable failure rates similar to those found on today's systems.

Several recent publications confirm that faults are common in high-performance computing systems. Therefore, further attention to the faults experienced by such computing systems is warranted. In this paper, we present a study of DRAM and SRAM faults in large high-performance computing systems. Our goal is to understand the factors that influence faults in production settings.

As high-end computing machines continue to grow in size, issues such as fault tolerance and reliability limit application scalability. Current techniques to ensure progress across faults, like checkpoint-restart, are increasingly problematic at these scales due to excessive overheads predicted to more than double an application's time to solution. Replicated computing techniques, particularly state machine replication, long used in distributed and mission critical systems, have been suggested as an alternative to checkpoint-restart. In this paper, we evaluate the viability of using state machine replication as the primary fault tolerance mechanism for upcoming exascale systems. We use a combination of modeling, empirical analysis, and simulation to study the costs and benefits of this approach in comparison to checkpoint/restart on a wide range of system parameters. These results, which cover different failure distributions, hardware mean time to failures, and I/O bandwidths, show that state machine replication is a potentially useful technique for meeting the fault tolerance demands of HPC applications on future exascale platforms.

Accurate fault detection is a key element of resilient computing. Syslogs provide key information regarding faults, and are found on nearly all computing systems. Discovering new fault types requires expert human effort, however, as no previous algorithm has been shown to localize faults in time and space with an operationally acceptable false positive rate. We present experiments on three weeks of syslogs from Sandia's 512-node "Spirit" Linux cluster, showing one algorithm that localizes 50% of faults with 75% precision, corresponding to an excellent false positive rate of 0.05%. The salient characteristics of this algorithm are (1) calculation of nodewise information entropy, and (2) encoding of word position. The key observation is that similar computers correctly executing similar work should produce similar logs.

We present Nodeinfo, an unsupervised algorithm for anomaly detection in system logs. We demonstrate Nodeinfo's effectiveness on data from four of the world's most powerful supercomputers: using logs representing over 746 million processor-hours, in which anomalous events called alerts were manually tagged for scoring, we aim to automatically identify the regions of the log containing those alerts. We formalize the alert detection task in these terms, describe how Nodeinfo uses the information entropy of message terms to identify alerts, and present an online version of this algorithm, which is now in production use. This is the first work to investigate alert detection on (several) publicly-available supercomputer system logs, thereby providing a reproducible performance baseline.

If we hope to automatically detect and diagnose failures in large-scale computer systems, we must study real deployed systems and the data they generate. Progress has been hampered by the inaccessibility of empirical data. This paper addresses that dearth by examining system logs from five supercomputers, with the aim of providing useful insight and direction for future research into the use of such logs. We present details about the systems, methods of log collection, and how alerts were identified; propose a simpler and more effective filtering algorithm; and define operational context to encompass the crucial information that we found to be currently missing from most logs. The machines we consider (and the number of processors) are: Blue Gene/L (131072), Red Storm (10880), Thunderbird (9024), Spirit (1028), and Liberty (512). This is the first study of raw system logs from multiple supercomputers.