We present here a report produced by a workshop on 'Addressing failures in exascale computing' held in Park City, Utah, 4-11 August 2012. The charter of this workshop was to establish a common taxonomy about resilience across all the levels in a computing system, discuss existing knowledge on resilience across the various hardware and software layers of an exascale system, and build on those results, examining potential solutions from both a hardware and software perspective and focusing on a combined approach.

The workshop brought together participants with expertise in applications, system software, and hardware; they came from industry, government, and academia, and their interests ranged from theory to implementation. The combination allowed broad and comprehensive discussions and led to this document, which summarizes and builds on those discussions.

/pdf/addressing-failures-in-exascale-computing-1b4vxg0uhl.pdf

Addressing failures in exascale computing

This paper provides an analysis of failures and their impact for Blue Waters, the Cray hybrid (CPU/GPU) supercomputer at the University of Illinois at Urbana-Champaign. The analysis is based on both manual failure reports and automatically generated event logs collected over 261 days. Results include i) a characterization of the root causes of single-node failures, ii) a direct assessment of the effectiveness of system-level fail over as well as memory, processor, network, GPU accelerator, and file system error resiliency, and iii) an analysis of system-wide outages. The major findings of this study are as follows. Hardware is not the main cause of system downtime. This is notwithstanding the fact that hardware-related failures are 42% of all failures. Failures caused by hardware were responsible for only 23% of the total repair time. These results are partially due to the fact that processor and memory protection mechanisms (x8 and x4 Chip kill, ECC, and parity) are able to handle a sustained rate of errors as high as 250 errors/h while providing a coverage of 99.997% out of a set of more than 1.5 million of analyzed errors. Only 28 multiple-bit errors bypassed the employed protection mechanisms. Software, on the other hand, was the largest contributor to the node repair hours (53%), despite being the cause of only 20% of the total number of failures. A total of 29 out of 39 system-wide outages involved the Lustre file system with 42% of them caused by the inadequacy of the automated fail over procedures.

Lessons Learned from the Analysis of System Failures at Petascale: The Case of Blue Waters

The increasing amount of data that needs to be collected and analyzed requires large-scale datacenter architectures that are naturally more susceptible to faults of single components. One way to offer consistent services on such unreliable systems are replicated state machines (RSMs). Yet, traditional RSM protocols cannot deliver the needed latency and request rates for future large-scale systems. In this paper, we propose a new set of protocols based on Remote Direct Memory Access (RDMA) primitives. To asses these mechanisms, we use a strongly consistent key-value store; the evaluation shows that our simple protocols improve RSM performance by more than an order of magnitude. Furthermore, we show that RDMA introduces various new options, such as log access management. Our protocols enable operators to fully utilize the new capabilities of the quickly growing number of RDMA-capable datacenter networks.

/pdf/dare-high-performance-state-machine-replication-on-rdma-14vr67hvfx.pdf

DARE: High-Performance State Machine Replication on RDMA Networks

A large percentage of computing capacity in today's large high-performance computing systems is wasted because of failures. Consequently current research is focusing on providing fault tolerance strategies that aim to minimize fault's effects on applications. By far the most popular technique is the checkpointrestart strategy. A complement to this classical approach is failure avoidance, by which the occurrence of a fault is predicted and preventive measures are taken. This requires a reliable prediction system to anticipate failures and their locations. Thus far, research in this field has used ideal predictors that were not implemented in real HPC systems. In this paper, we merge signal analysis concepts with data mining techniques to extend the ELSA (Event Log Signal Analyzer) toolkit and offer an adaptive and more efficient prediction module. Our goal is to provide models that characterize the normal behavior of a system and the way faults affect it. Being able to detect deviations from normality quickly is the foundation of accurate fault prediction. However, this is challenging because component failure dynamics are heterogeneous in space and time. To this end, a large part of the paper is focused on a detailed analysis of the prediction method, by applying it to two large-scale systems and by investigating the characteristics and bottlenecks of each step of the prediction process. Furthermore, we analyze the prediction's precision and recall impact on current checkpointing strategies and highlight future improvements and directions for research in this field.

/pdf/fault-prediction-under-the-microscope-a-closer-look-into-hpc-2cb61zwr5y.pdf

Fault prediction under the microscope: a closer look into HPC systems

This timely text presents a comprehensive overview of fault tolerance techniques for high-performance computing (HPC). The text opens with a detailed introduction to the concepts of checkpoint protocols and scheduling algorithms, prediction, replication, silent error detection and correction, together with some application-specific techniques such as ABFT. Emphasis is placed on analytical performance models. This is then followed by a review of general-purpose techniques, including several checkpoint and rollback recovery protocols. Relevant execution scenarios are also evaluated and compared through quantitative models. Features: provides a survey of resilience methods and performance models; examines the various sources for errors and faults in large-scale systems; reviews the spectrum of techniques that can be applied to design a fault-tolerant MPI; investigates different approaches to replication; discusses the challenge of energy consumption of fault-tolerance methods in extreme-scale systems.

/pdf/fault-tolerance-techniques-for-high-performance-computing-28z1e6yekc.pdf

Fault-Tolerance Techniques for High-Performance Computing

As supercomputers and clusters increase in size and complexity, system failures are inevitable. Different hardware components (such as memory, disk, or network) of such systems can have different failure rates. Prior works assume failures equally affect an application, whereas our goal is to provide failure models for applications that reflect their specific component usage. This is challenging because component failure dynamics are heterogeneous in space and time. To this end, we study 5 years of system logs from a production high-performance computing system and model hardware failures involving processors, memory, storage and network components. We model each component and construct integrated failure models given the component usage of common supercomputing applications. We show that these application-centric models provide more accurate reliability estimates compared to general models, which improves the efficacy of fault-tolerant algorithms. In particular, we demonstrate how applications can tune their checkpointing strategies to the tailored model.

/pdf/modeling-and-tolerating-heterogeneous-failures-in-large-8dh036xc6l.pdf

Dan LaPine

Papers

Modeling and tolerating heterogeneous failures in large parallel systems