Application　identification　plays　an　essential　role　in　network　management　such　as　intrusion　detection　and　security　monitoring.　But　the　continuous　growth　of　bandwidth　and　massive　amount　of　packets　pose　serious　challenges　for　efficacious　and　accurate　application　identification.　In　this　paper,　we　develop　a　new　method　to　reduce　the　number　of　packets　being　processed　while　achieving　the　goal　of　accurate　P2P　and　VoIP　application　identification.　Firstly,　we　employ　the　Bi-flow　model　to　aggregate　traffic　packets　into　Bi-flow,　which　can　capture　the　exchange　behavior　characteristics　between　different　terminals.　Then　we　employ　the　signature　of　Packet　Size　Distribution　(PSD)　to　capture　flow　dynamics,　which　is　defined　as　the　payload　length　distribution　probability　of　the　packets　in　one　Bi-flow.　Secondly,　we　collect　PSD　of　several　different　P2P　and　VoIP　applications　and　the　analysis　results　show　that　PSD　of　different　applications　are　different　with　each　other,　which　can　be　used　as　features　to　perform　traffic　identification.　We　also　find　the　PSD　characteristics　of　one　Bi-flow　can　be　captured　by　its　first　few　packets,　which　demonstrate　our　methods　can　identify　the　Bi-flow,quickly　after　its　establishment.　We　employ　the　Renyi　cross　entropy　to　perform　identification　by　calculating　the　similarity　between　PSD　of　the　Bi-flow　being　identified　and　that　of　specific　application.　If　the　similarity　is　higher　than　a　selected　threshold,　the　Bi-flow　being　identified　is　classified　to　the　specific　application.　Finally,　as　the　PSD　is　a　type　of　probability　feature　which　is　not　sensitive　to　packet　lose,　we　integrate　the　Poisson　sampling　method　into　our　framework　to　process　the　massive　data　in　backbone　networks.　Experimental　results　using　the　artificial　and　actual　traces　collected　from　monitoring　platform　in　the　Northwest　Center　of　CERNET　(China　Education　and　Research　Network)　verify　the　accuracy　and　robustness　of　our　method.　(C)　2015　Elsevier　B.V.　All　rights　reserved.