This document discusses using machine learning algorithms for predictive performance modeling of IT systems. It explains that ML can be used to predict quality of service metrics like response time and server utilization for different conditions, such as increased user load or hardware configurations, based on past production and test data. This helps with data-driven decision making for hardware procurement and effective utilization, and can reduce the cost and time of performance testing by complementing benchmarking and application tuning. The key is having sufficient historical data to train accurate models, with various techniques available along a cost-accuracy spectrum, from linear projections to simulation to machine learning.