In the context of optimizing Spark performance for a large-scale machine learning project, which technique is used to store intermediate data in memory, thereby speeding up iterative algorithms and reducing disk I/O?