arXiv subjects

Ozalp Babaoglu

Publications and source records attributed to Ozalp Babaoglu.

10 recordsLinked to original sources

arXiv2020

A Machine Learning Approach to Online Fault Classification in HPC Systems

As High-Performance Computing (HPC) systems strive towards the exascale goal, failure rates both at the hardware and software levels will increase significantly. Thus, detecting and classifying faults in HPC systems as they occur and initiating corrective actions before they can transform into failures becomes essential for continued operation. Central to this objective is fault injection, which is the deliberate triggering of faults in a system so as to observe their behavior in a controlled environment. In this paper, we propose a fault classification method for HPC systems based on machine learning. The novelty of our approach rests with the fact that it can be operated on streamed data in an online manner, thus opening the possibility to devise and enact control actions on the target system in real-time. We introduce a high-level, easy-to-use fault injection tool called FINJ, with a focus on the management of complex experiments. In order to train and evaluate our machine learning classifiers, we inject faults to an in-house experimental HPC system using FINJ, and generate a fault dataset which we describe extensively. Both FINJ and the dataset are publicly available to facilitate resiliency research in the HPC systems field. Experimental results demonstrate that our approach allows almost perfect classification accuracy to be reached for different fault types with low computational overhead and minimal delay.

cs.DC↗

arXiv2019

Online Fault Classification in HPC Systems through Machine Learning

As High-Performance Computing (HPC) systems strive towards the exascale goal, studies suggest that they will experience excessive failure rates. For this reason, detecting and classifying faults in HPC systems as they occur and initiating corrective actions before they can transform into failures will be essential for continued operation. In this paper, we propose a fault classification method for HPC systems based on machine learning that has been designed specifically to operate with live streamed data. We cast the problem and its solution within realistic operating constraints of online use. Our results show that almost perfect classification accuracy can be reached for different fault types with low computational overhead and minimal delay. We have based our study on a local dataset, which we make publicly available, that was acquired by injecting faults to an in-house experimental HPC system.

cs.DC↗

arXiv2018

FINJ: A Fault Injection Tool for HPC Systems

We present FINJ, a high-level fault injection tool for High-Performance Computing (HPC) systems, with a focus on the management of complex experiments. FINJ provides support for custom workloads and allows generation of anomalous conditions through the use of fault-triggering executable programs. FINJ can also be integrated seamlessly with most other lower-level fault injection tools, allowing users to create and monitor a variety of highly-complex and diverse fault conditions in HPC systems that would be difficult to recreate in practice. FINJ is suitable for experiments involving many, potentially interacting nodes, making it a very versatile design and evaluation tool.

cs.DC↗

arXiv2016

Towards Operator-less Data Centers Through Data-Driven, Predictive, Proactive Autonomics

Continued reliance on human operators for managing data centers is a major impediment for them from ever reaching extreme dimensions. Large computer systems in general, and data centers in particular, will ultimately be managed using predictive computational and executable models obtained through data-science tools, and at that point, the intervention of humans will be limited to setting high-level goals and policies rather than performing low-level operations. Data-driven autonomics, where management and control are based on holistic predictive models that are built and updated using live data, opens one possible path towards limiting the role of operators in data centers. In this paper, we present a data-science study of a public Google dataset collected in a 12K-node cluster with the goal of building and evaluating predictive models for node failures. Our results support the practicality of a data-driven approach by showing the effectiveness of predictive models based on data found in typical data center logs. We use BigQuery, the big data SQL platform from the Google Cloud suite, to process massive amounts of data and generate a rich feature set characterizing node state over time. We describe how an ensemble classifier can be built out of many Random Forest classifiers each trained on these features, to predict if nodes will fail in a future 24-hour window. Our evaluation reveals that if we limit false positive rates to 5%, we can achieve true positive rates between 27% and 88% with precision varying between 50% and 72%.This level of performance allows us to recover large fraction of jobs' executions (by redirecting them to other nodes when a failure of the present node is predicted) that would otherwise have been wasted due to failures. [...]

cs.DC↗

arXiv2016

Power Consumption Modeling and Prediction in a Hybrid CPU-GPU-MIC Supercomputer (preliminary version)

Power consumption is a major obstacle for High Performance Computing (HPC) systems in their quest towards the holy grail of ExaFLOP performance. Significant advances in power efficiency have to be made before this goal can be attained and accurate modeling is an essential step towards power efficiency by optimizing system operating parameters to match dynamic energy needs. In this paper we present a study of power consumption by jobs in Eurora, a hybrid CPU-GPU-MIC system installed at the largest Italian data center. Using data from a dedicated monitoring framework, we build a data-driven model of power consumption for each user in the system and use it to predict the power requirements of future jobs. We are able to achieve good prediction results for over 80% of the users in the system. For the remaining users, we identify possible reasons why prediction performance is not as good. Possible applications for our predictive modeling results include scheduling optimization, power-aware billing and system-scale power modeling. All the scripts used for the study have been made available on GitHub.

cs.DC↗

arXiv2016

Predicting System-level Power for a Hybrid Supercomputer

For current High Performance Computing systems to scale towards the holy grail of ExaFLOP performance, their power consumption has to be reduced by at least one order of magnitude. This goal can be achieved only through a combination of hardware and software advances. Being able to model and accurately predict the power consumption of large computational systems is necessary for software-level innovations such as proactive and power-aware scheduling, resource allocation and fault tolerance techniques. In this paper we present a 2-layer model of power consumption for a hybrid supercomputer (which held the top spot of the Green500 list on July 2013) that combines CPU, GPU and MIC technologies to achieve higher energy efficiency. Our model takes as input workload information - the number and location of resources that are used by each job at a certain time - and calculates the resulting system-level power consumption. When jobs are submitted to the system, the workload configuration can be foreseen based on the scheduler policies, and our model can then be applied to predict the ensuing system-level power consumption. Additionally, alternative workload configurations can be evaluated from a power perspective and more efficient ones can be selected. Applications of the model include not only power-aware scheduling but also prediction of anomalous behavior.

cs.DC↗

arXiv2015

A Holistic Approach to Log Data Analysis in High-Performance Computing Systems: The Case of IBM Blue Gene/Q

The complexity and cost of managing high-performance computing infrastructures are on the rise. Automating management and repair through predictive models to minimize human interventions is an attempt to increase system availability and contain these costs. Building predictive models that are accurate enough to be useful in automatic management cannot be based on restricted log data from subsystems but requires a holistic approach to data analysis from disparate sources. Here we provide a detailed multi-scale characterization study based on four datasets reporting power consumption, temperature, workload, and hardware/software events for an IBM Blue Gene/Q installation. We show that the system runs a rich parallel workload, with low correlation among its components in terms of temperature and power, but higher correlation in terms of events. As expected, power and temperature correlate strongly, while events display negative correlations with load and power. Power and workload show moderate correlations, and only at the scale of components. The aim of the study is a systematic, integrated characterization of the computing infrastructure and discovery of correlation sources and levels to serve as basis for future predictive modeling efforts.

cs.DC↗

arXiv2015

A Big Data Analyzer for Large Trace Logs

Current generation of Internet-based services are typically hosted on large data centers that take the form of warehouse-size structures housing tens of thousands of servers. Continued availability of a modern data center is the result of a complex orchestration among many internal and external actors including computing hardware, multiple layers of intricate software, networking and storage devices, electrical power and cooling plants. During the course of their operation, many of these components produce large amounts of data in the form of event and error logs that are essential not only for identifying and resolving problems but also for improving data center efficiency and management. Most of these activities would benefit significantly from data analytics techniques to exploit hidden statistical patterns and correlations that may be present in the data. The sheer volume of data to be analyzed makes uncovering these correlations and patterns a challenging task. This paper presents BiDAl, a prototype Java tool for log-data analysis that incorporates several Big Data technologies in order to simplify the task of extracting information from data traces produced by large clusters and server farms. BiDAl provides the user with several analysis languages (SQL, R and Hadoop MapReduce) and storage backends (HDFS and SQLite) that can be freely mixed and matched so that a custom tool for a specific task can be easily constructed. BiDAl has a modular architecture so that it can be extended with other backends and analysis languages in the future. In this paper we present the design of BiDAl and describe our experience using it to analyze publicly-available traces from Google data clusters, with the goal of building a realistic model of a complex data center.

cs.DC↗

arXiv2015

Towards Data-Driven Autonomics in Data Centers

Continued reliance on human operators for managing data centers is a major impediment for them from ever reaching extreme dimensions. Large computer systems in general, and data centers in particular, will ultimately be managed using predictive computational and executable models obtained through data-science tools, and at that point, the intervention of humans will be limited to setting high-level goals and policies rather than performing low-level operations. Data-driven autonomics, where management and control are based on holistic predictive models that are built and updated using generated data, opens one possible path towards limiting the role of operators in data centers. In this paper, we present a data-science study of a public Google dataset collected in a 12K-node cluster with the goal of building and evaluating a predictive model for node failures. We use BigQuery, the big data SQL platform from the Google Cloud suite, to process massive amounts of data and generate a rich feature set characterizing machine state over time. We describe how an ensemble classifier can be built out of many Random Forest classifiers each trained on these features, to predict if machines will fail in a future 24-hour window. Our evaluation reveals that if we limit false positive rates to 5%, we can achieve true positive rates between 27% and 88% with precision varying between 50% and 72%. We discuss the practicality of including our predictive model as the central component of a data-driven autonomic manager and operating it on-line with live data streams (rather than off-line on data logs). All of the scripts used for BigQuery and classification analyses are publicly available from the authors' website.

cs.DC↗

arXiv2014

BiDAl: Big Data Analyzer for Cluster Traces

Modern data centers that provide Internet-scale services are stadium-size structures housing tens of thousands of heterogeneous devices (server clusters, networking equipment, power and cooling infrastructures) that must operate continuously and reliably. As part of their operation, these devices produce large amounts of data in the form of event and error logs that are essential not only for identifying problems but also for improving data center efficiency and management. These activities employ data analytics and often exploit hidden statistical patterns and correlations among different factors present in the data. Uncovering these patterns and correlations is challenging due to the sheer volume of data to be analyzed. This paper presents BiDAl, a prototype "log-data analysis framework" that incorporates various Big Data technologies to simplify the analysis of data traces from large clusters. BiDAl is written in Java with a modular and extensible architecture so that different storage backends (currently, HDFS and SQLite are supported), as well as different analysis languages (current implementation supports SQL, R and Hadoop MapReduce) can be easily selected as appropriate. We present the design of BiDAl and describe our experience using it to analyze several public traces of Google data clusters for building a simulation model capable of reproducing observed behavior.

cs.DC↗