As　a　service-oriented　parallel　distributed　computing　paradigm,　cloud　computing　can　tackle　large-scale　computing　problem　by　cloud　resources.　A　challenge　to　optimize　cloud　resource　utilization　is　more　efficient　scheduling　users＇　requests　(workflows).　However,　most　of　algorithms　assume　that　cloud　resources＇　performance　is　always　fixed,　which　is　impractical　due　to　the　uncertainty　during　the　task　execution.　In　this　paper,　workflow　scheduling　considering　the　performance　variation　of　cloud　resources　is　studied　aiming　to　minimize　the　makespan,　which　is　formulated　as　a　Markov　Decision　Process.　And,　a　dynamic　workflow　scheduling　approach　based　on　deep　reinforcement　learning　(RLWS)　is　proposed.　In　this　approach,　a　complete　solution　is　as　the　input,　and　neural　network　parameters　are　learned　by　iteratively　local　re-scheduling　to　optimize　the　solution.　Actor　critic　in　deep　reinforcement　learning　is　designed　to　train　the　neural　network　parameters　by　self-learning　procedure.　Experiment　results　confirm　that　the　proposed　algorithm　can　efficiently　shorten　the　makespan.