To　minimize　Flow　Completion　Time　(FCT),　existing　flow　scheduling　schemes　assume　prior　knowledge　of　accurate　per-flow　information,　eg,　flow　sizes　or　deadlines,　to　achieve　superior　performance.　In　practice,　it　is　hard　to　get　accurate　per-flow　information,　especially　in　multi-tenant　cloud　environments.　Rather　than　such　unrealistic　assumption　(using　accurate　per-flow　information),　this　paper　proposes　a　flow　size　estimation　mechanism　(called　LFE),　which　uses　machine　learning　algorithms　to　learn　and　explore　the　flow　characteristics　or　patterns　from　historical　data.　LFE　can　estimate　the　flow　size　rapidly　without　accurate　per-flow　information.　To　evaluate　the　impact　of　flow　size　estimation　on　flow　scheduling　performance,　we　implement　LFE　in　a　flow-level　simulator　and　test　its　performance　with　KMeans　and　PageRank　workload,　respectively.　Compared　with　FLUX,　the　average　FCT　reduces　13%　at　90%　load.　The　results　show　that　LFE　has　a　better　flow　size　prediction　accuracy　and　can　improve　the　flow　scheduling　performance.