Efficient Cross-Rack Data Updates Based on Intelligent Dynamic Data Layout Using Machine-learning

In large-scale storage systems, erasure-coded data updates often cause performance issues due to high inter-rack traffic. This paper introduces CRUD, an efficient method to reduce inter-rack traffic. CRUD uses combinatorial mathematics to create a pairwise balanced design (PBD), distributing stripes across racks to exploit data update locality. It applies a machine-learning model to classify data blocks into repeated and non-repeated updates. CRUD considers update categories, strategies, block numbers, and rack involvement to optimize stripe distribution, dynamically adjusting it to further reduce traffic. Experiments show CRUD cuts inter-rack traffic by up to 42.31% and saves 31.95% in update time compared to the best current method.