MACHINE LEARNING FOR PERFORMANCE IMPROVEMENT ON SSD

• Main Author: Soobhan Zulkifli, Kahfi
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79447
• Institution: Institut Teknologi Bandung

Currently, SSD storage systems have unstable performance caused by internal processes which can be seen in the "tail" phenomenon or elongation of the tail in CDF latency. This research uses a machine learning approach to reduce the lengthening of CDF latency. Data is replicated across a number of SSDs and machine learning determines which SSDs are considered fast SSDs from SSD user requests. This research takes I/O data from AliBaba, Microsoft, and Tencent. The data was run on the FEMU SSD emulator to get latency data and this data was used to train a machine learning model using AutoML with the auto-sklearn tool. With machine learning, it was found that performance on SSDs could be improved as evidenced by the movement of the CDF latency graph to the left. The best models obtained are gradient boosting models, random trees, and extra trees because these models are ensemble models which are arrangements consisting of various different models.