A New P2P Network Routing Algorithm Based on ISODATA Clustering Topology

Transcription

1 Avalable onlne at Proeda Engneerng 5 (20) Advaned n Control Engneerngand Inforaton Sene A ew P2P etwork Routng Algorth Based on ISODATA Clusterng Topology Y Ma a, Zhenhua Tan a, Guran Chang a, Xuey Wang a, a* a Software College, ortheastern Unversty, Shenyang Cty, Laonng Provne,Chna Abstrat Most P2P applatons use the routng algorth that seletng the neghbor nodes at rando. These routng algorths nrease the routng hops. To get better routng effeny, a new routng algorth naed RIDC was presented n ths paper. It dynaally erges nodes nto dfferent lusters n a taxonoy herarhy, and organzes the lusters nto routng overlays. By ths algorth the network perforane s greatly enhaned. Prelnary evaluaton shows that RIDC aheves a good onvergene on a large sale of nodes. 20 Publshed by Elsever Ltd. Open aess under CC BY-C-D lense. Seleton and/or peer-revew under responsblty of [CEIS 20] Keywords P2P; ISODATA; herarhy aggloeratve lusterng; routng algorth. Introduton Most lassal P2P network odals, for exaple, Freenet[], Pastry[2], Kadela[3], Chord[4], Tapestry[5] and CA[6], are based on Dstrbuted Hash Table(DHT). These algorths are effetve opared to other unstrutured P2P routng algorths. Soe algorth that every node reords routng essage of all other nodes s even ore effetve, but not approprate for large sale of nodes wth ore start-up te and bandwdth. Ths paper presents a new routng algorth based on ISODATA Clusterng Topology (RIDC) for strutured P2P overlay network. Aordng to the ounaton hstory, lusters are erged by ISODATA (Iteratve Self-Organzng Data Analyss Tehnque) [7] ethod. Eah luster has a SubSuper- ode, by whh the nodes n the sae luster an route dretly. Meanwhle, the lusters are organzed as * Y Ma. Tel.: E-al address: yron_y@63.o Publshed by Elsever Ltd. do:0.06/.proeng Open aess under CC BY-C-D lense.

2 Y Ma et al. / Proeda Engneerng 5 (20) a tree by the slarty of eah other n aggloeratve lusterng algorth [8]. So the whole network has a good routng effeny. 2. ISODATA Clusterng The ISODATA algorth has soe further refneents by splttng and ergng of lusters than other algorths. Clusters are erged f ether the nuber of ebers n a luster s less than a ertan threshold or f the enters of two lusters are loser than a ertan threshold. Clusters are splt nto two dfferent lusters f the luster standard devaton exeeds a predefned value and the nuber of ebers s twe the threshold for the nu nuber of ebers. The new P2P network topology n ths paper s strutured by several lusters whh were oposed by a tree herarhy. All the nodes are subordnated to the top luster, and every luster an be splt nto sub lusters. The root node of the tree s alled Super-ode, and the root node of sub luster s SubSuper- ode. In the sae luster, every node s equal whether noral node or root node. In the tree strutured herarhy topology, every root node of sub luster has routng nforaton of other nodes n the sae luster. Defnton of Counaton Hstory To opute the slarty of nodes n P2P network, every node keeps a log fle that shows the ounaton hstory. S, λ and η s data quantty, tes and proporton of suessful ounatons of node. We an defne a 3-denson vetor β(), whh shows aount of ounaton, rate of suessful ounaton, and quantty of ounaton data. = x S S + y λ λ + z η η () ISODATA Clusterng Algorth In P2P In ISODATA lusterng algorth, lusters wll be erged f ether the nuber of ebers n a luster are less than a ertan threshold or f the enters of two lusters are loser than a ertan threshold. Meanwhle, lusters an also splt nto two dfferent lusters f the luster standard devaton exeeds a predefned value and the nuber of ebers s twe the threshold for the nu nuber of ebers. The prnple s as follow. [Algorth ] ISODATA Clusterng Algorth K s the antpant aount of lusters. θ s the nu nuber of a luster. θ S s the paraeter of standard devaton. θ C s the paraeter of ergng. L s the axu nuber an be erged by one teraton. I s the teratve nuber. Step: Intalze seeds to be the luster enters,, =,2,,. Step2: f y < y =,2,..., then y Γ In ths way, all the nodes are assgned to the lusters. Step3: For eah luster Γ f < θ then delete Γ, =-. Step4: = y, y =, 2,..., y Γ Step5: Calulate the average dstaneδ between node y and n luster δ = y, =, 2,..., y Γ Step6: Calulate the average dstaneδ of all nodes. δ = δ, =, 2,..., = Γ. (2) (3) (4)

3 2968 Y Ma et al. / Proeda Engneerng 5 (20) Step7: f <=K/2, goto step 8 f >=2K, goto step else goto step 8 Step8: For eah luster Γ, alulate the standard devaton T [, 2 d ] 2 = ( y k ) (6) y k Γ =,..., (5) y k s the th feature of the kth node n luster, s the th feature of, s the standard devaton of th feature n luster, d s the denson of y. Step9: For eah luster Γ, alulate the axu feature ax, =,2,...,. Step0: For eah ax > θ s f ( > and >2(θ +)) or(<=k/2) + then splt Γ nto two dfferent lusters, whh luster enters are and, =+. + = + k, 0 < k <= (7) = k, 0 < k <= (8) Step: For eah and =,2,..., = (9) = +,..., Step2: For eah < θc, sort l of the < 2 2 <... < ll Step3: Merge the sallest and, =- l = [ + ] + (0) Step4: I=I- If I==0 then return else goto step 2 Usng ISODATA lusterng algorth, the nodes wll be lassfed nto several lusters, whh are not too large or too sall. The te effeny of ISODATA algorth s O ( 2 ). 3. Construton of Tree Strutured Herarhy Topology The lusters obtaned by ISODATA wll be onstruted nto herarhy topology by group-average alloeratve lusterng algorth (GAAC). Group-Average Alloeratve Clusterng In P2P Group-average aggloeratve lusterng algorth evaluates luster qualty based on all slartes between douents, thus avodng the ptfalls of the sngle-lnk and oplete-lnk rtera, whh equate luster slarty wth the slarty of a sngle par of douents. GAAC alulates the average slarty of all pars of douents, nludng pars fro the sae luster. As every node s a sub luster of ts upper luster exept the root node, the nodes an be lay out as a tree. We use a botto-up lusterng algorth as follow. step: Gven: a set X ={x,,x n } of obets and a funton s step2: for = to n do ={x } step3: C={,... n } step4: whle C > step5: ( n, n 2 ) = arg ax (, ) s(, ) u v u v

4 Y Ma et al. / Proeda Engneerng 5 (20) step6: ; = n U n 2 C = C \ { n, n 2 } U { } In ths paper, the set X s obtaned by ISODATA. Group-average slarty S( ) an be alulated as follow. () S ( ) = s ( x, y) ( ) x x y The group-average slarty of the unted luster an be alulated as follow. ( S ( ) + S ( )) ( S ( ) + S ( )) ( + ) (2) S ( U ) = ( + )( + ) When GAAC was fnshed, we got a tree strutured topology. In eah luster there s a root node naed SubSuper-ode that provdes a ost relable perforane. When the lusterng algorth s done, every node n the sae luster wll be regstered by the SubSuper- ode so that t an route dretly to other nodes. odes Jon and Ext When a new node ons the P2P network, t wll not be put nto any luster untl the traff exeeds a ertan threshold, then t wll be put nto the luster that t ounated wth ost frequently.when a node s leavng the network, there are two ases to dsuss. If a leaf node s leavng, we an delete t fro the network dretly, and nfor ts hgher SubSuper-ode. If a SubSuper-ode s leavng, we should use ts bakup root node to take plae of t and rebuld the luster. 4. Routng Algorth [Algorth 2] Routng Algorth Input: P2P syste, key_d of the destnaton reourse, node p that starts the route and a routng essage sg. Output: node p_dst who s the last node n the route that has the key_d resoure. step: Searh for the SubSuper-ode n the luster. ode p sends sg to ts own luster root node C. step2: If C exsts, C searhes n ts luster.if the there s a node owns the key_d, then go to step 5, and f not, then go to step 3.If C doesn t exst, that eans the P2P syste s busy lusterng and updatng the new route table. So p has to wat untl ts new SubSuper-ode C s seleted. step3: C sends essage to ts hgher SubSuper-ode to fnd the owner of key_d. If there s a responder naed C fro other luster, then go to step 4; and f not, then go to step 6. step4: C searhes n ts own luster.if the there s a node owns the key_d, then go to step 5; and f not, then go to step 6. step5: Return p_dst. step6: Return null. Aordng to ths algorth, f the routng s suessful, there s a good perforane of routng hops wth the oplexty of O(log).The lusterng s not runnng frequently, t wll start lusterng and seletng new luster enter nodes only neessary. As desrbed, the te effeny of RIDC algorth s O(log). In ost of te, we wll get several lusters by ISODATA, and then these lusters wll be strutured as a tree, so the searhng te wll be the heght of the tree, whh s O(). 5. Perforanes We opare the routng hops wth hord [4] and RIDC n a sulate syste. Whle the nodes are nreasng fro 000 to 0000, we reord the routng hops, as shown n Fgure. Obvously, wth new network topology of RIDC, the routng hops rean a low level. Ths s the oparson n a steady network, whh eans there s no node on or reet fro the P2P network. In RIDC, before nodes are

5 2970 Y Ma et al. / Proeda Engneerng 5 (20) strutured n a tree topology, they are erged nto several lusters by ISODATA, so the heght of the tree s uh lower than O (log ), t s between O () and O (log ). 0 8 Hops Chord RIDC 0 K 2K 3K 4K 5K 6K 7K 8K 9K 0K odes Fg.. Coparson for routng hops Aordng to the nodes leavng algorth, upper SubSuper-odes wll be nfored fro botto to top f a noral node s leavng the P2P syste. The essage nuber s equal to the heght of the tree, so the te oplexty of nodes leavng n RIDC s O (log ). If a SubSuper-ode s leavng, the bakup root node n the sae luster wll take plae of t and update the route table fro botto to top. So the oplexty of deleton n RIDC s stll O (log ). Message nuber of nodes leavng s ore than nodes on beause soe luster ay be reorganzed as the SubSuper-ode s hanged. Copared wth Chord, RIDC s effent, and ths eans lusterng by a tree strutured herarhy topology s better than DHT. 6. Conlusons In onluson, the routng algorth RIDC has advantages as follows. () Desgn an prove herarhy aggloeratve lusterng algorth based on ISODATA, by whh the P2P network are organzed nto tree strutured herarhy topology. (2) Optze the rout table aordng to the lusters, so we an nrease the perforane n routng effeny. (3) Wth the te oplexty of O (log ), RIDC shows a good stablty for edu-large sze P2P syste. Referenes [] I. Clarke, T Hong,S. Mller. Protetng free expresson onlne wth Freenet. IEEE Internet Coputng,6():pp.40-49,2002 [2] A. Rowstron and P. Drushel, Pastry: Salable, dstrbuted obet loaton and routng for large-sale peer-to-peer systes, n Proeedngs of the Mddleware, 200. [3] Petar Mayounkov,Davd Mazeres. Kadel: A Peer-to-Peer Inforaton Syste Based on the XOR Metr. [4] I. Stoa, R. Morrs, D. Karger, M. F. Kaashoek, and H. Balakrshnan, Chord: A salable peer-to-peer lookup protool for nternet applatons, IEEE/ACM Transatons on etworkng, vol., no., pp. 7 32, [5] B. Y. Zhao, L. Huang, J. Strblng, S. C. Rhea, A. D. Joseph, and J. D. Kubatowz, Tapestry: A reslent global-sale overlay for serve deployent, IEEE Journal on Seleted Areas n Counatons, vol. 22, no., pp. 4 53, January [6] S. Ratnasay, P. Frans, M. Handle. A salable ontent-addressable network. ACM SIGCOMM, San Dego, pp.6-72,200. [7] L Xu-lang, Ln We, Xu We-dong, Pan Yu-long, Sun Wen-yan; Applyng the Fuzzy Clusterng Analyss of ISODATA to the Classfaton of Caouflage Effetveness[J];Ata Araentar; [8] Ward, J. H. Herarhal groupngs to optze an obetve funton. Journal of the Aeran Statstal Assoaton, pp.58,