use Marshal instead of YAML when interoperability is not important

YAML is slow. Even in research dataset, the data can be pretty huge, if interoperability is not that important, use Marshal.dump/load to manipulate the data. It is very fast, but the file is not readable by both people and other languages. I've some experience waiting for the YAML dumping some 50000 elements hash for more than three hours. Ater a quick marshalling, better let the YAML run again slowly when sleeping if you also want interoperability.