MACHINE LEARNING DRIFT DETECTOR FOR DRIFT MITIGATION IN I/O STREAM WORKLOAD TO CUT TAIL LATENCY

• Main Author: Ayu Putri Irawan, Maharani
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/82433
• Institution: Institut Teknologi Bandung

In production systems, it is common to use more than one SSD for redundant storage. These redundant SSDs can help manage incoming I/O requests to avoid overloading an SSD that is conducting its internal management tasks, such as garbage collection, wear-leveling, and write amplification, by utilizing I/O admission control. Machine learning algorithms are currently widely used in various production systems that encounter continuous data, where model drift can occur. Machine learning-based I/O admission control, as one of the applications in production systems, also encounters model drift, which consists of concept drift and covariate shift. Both types of drift can be mitigated by detecting their occurrence using certain conditions, such as model accuracy and statistics-based algorithms, and then retraining the model to adapt to the characteristics of the current data. This study applies two schemes and introduces one innovative machine learning-based scheme to mitigate drift. Those algorithms are, respectively, model reuse, model reweight, and model-based drift detector.