پريسا علائي

High-performance computing (HPC) systems play a pivotal role in scientific projects due to their high processing power and ability to execute complex computations. However, anomaly detection in these systems has become a critical challenge, as anomalies can negatively impact the performance, resilience, and availability of these infrastructures. This challenge is primarily attributed to the lack of real and labeled data for accurately analyzing system behavior. This research proposes a deep learning model based on an autoencoder architecture with LSTM layers for analyzing temporal and complex data. The proposed model, utilizing three encoder layers and three decoder layers, has achieved an anomaly detection accuracy improvement of up to 91%. By leveraging its ability to learn complex temporal patterns, the model demonstrates optimized performance in anomaly detection. To address the limitations caused by the scarcity of real data, synthetic data generation methods have been developed. These methods include aggregating information from similar nodes and augmenting data through inter-row relationship analysis. Results show that these approaches significantly enhance data quality and realism, thereby improving the performance of the anomaly detection system. Other contributions of this research include designing a comprehensive eva‎luation process and comparing the proposed model with existing methods. The results highlight the model's superior performance in metrics such as precision, recall, and error detection rate. By focusing on improving the efficiency, resilience, and effectiveness of HPC systems, this research takes a significant step toward developing advanced solutions for monitoring and managing these infrastructures.

تشخيص ناهنجاري , سيستم‌هاي محاسباتي با كارآمدي بالا , خودرمزگذار , يادگيري عميق , توليد داده , مصرف CPU

Anomaly detection , High-Performance computing system , Deep learning , Autoencoder , CPU consumption , Data Generation

parisa alaie

Dr. Marzieh Malekimajd