Artificial intelligence driven anomaly detection for big data systems

Abstract

The main goal of this thesis is to contribute to the research on automated performance anomaly detection and interference prediction by implementing Artificial Intelligence (AI) solutions for complex distributed systems, especially for Big Data platforms within cloud computing environments. The late detection and manual resolutions of performance anomalies and system interference in Big Data systems may lead to performance violations and financial penalties. Motivated by this issue, we propose AI-based methodologies for anomaly detection and interference prediction tailored to Big Data and containerized batch platforms to better analyze system performance and effectively utilize computing resources within cloud environments. Therefore, new precise and efficient performance management methods are the key to handling performance anomalies and interference impacts to improve the efficiency of data center resources.

The first part of this thesis contributes to performance anomaly detection for in-memory Big Data platforms. We examine the performance of Big Data platforms and justify our choice of selecting the in-memory Apache Spark platform. An artificial neural network-driven methodology is proposed to detect and classify performance anomalies for batch workloads based on the RDD characteristics and operating system monitoring metrics. Our method is evaluated against other popular machine learning algorithms (ML), as well as against four different monitoring datasets. The results prove that our proposed method outperforms other ML methods, typically achieving 98–99% F-scores. Moreover, we prove that a random start instant, a random duration, and overlapped anomalies do not significantly impact the performance of our proposed methodology.

The second contribution addresses the challenge of anomaly identification within an in-memory streaming Big Data platform by investigating agile hybrid learning techniques. We develop TRACK (neural neTwoRk Anomaly deteCtion in sparK) and TRACK-Plus, two methods to efficiently train a class of machine learning models for performance anomaly detection using a fixed number of experiments. Our model revolves around using artificial neural networks with Bayesian Optimization (BO) to find the optimal training dataset size and configuration parameters to efficiently train the anomaly detection model to achieve high accuracy. The objective is to accelerate the search process for finding the size of the training dataset, optimizing neural network configurations, and improving the performance of anomaly classification. A validation based on several datasets from a real Apache Spark Streaming system is performed, demonstrating that the proposed methodology can efficiently identify performance anomalies, near-optimal configuration parameters, and a near-optimal training dataset size while reducing the number of experiments up to 75% compared with naïve anomaly detection training.

The last contribution overcomes the challenges of predicting completion time of containerized batch jobs and proactively avoiding performance interference by introducing an automated prediction solution to estimate interference among colocated batch jobs within the same computing environment. An AI-driven model is implemented to predict the interference among batch jobs before it occurs within system. Our interference detection model can alleviate and estimate the task slowdown affected by the interference. This model assists the system operators in making an accurate decision to optimize job placement. Our model is agnostic to the business logic internal to each job. Instead, it is learned from system performance data by applying artificial neural networks to establish the completion time prediction of batch jobs within the cloud environments. We compare our model with three other baseline models (queueing-theoretic model, operational analysis, and an empirical method) on historical measurements of job completion time and CPU run-queue size (i.e., the number of active threads in the system). The proposed model captures multithreading, operating system scheduling, sleeping time, and job priorities. A validation based on 4500 experiments based on the DaCapo benchmarking suite was carried out, confirming the predictive efficiency and capabilities of the proposed model by achieving up to 10% MAPE compared with the other models.