Efficient Subscription Management in Content-based Networks

Transcription

1 Effiient Susription Mngement in Content-sed Networks Rphël Chnd, Psl A. Feler Institut EURECOM Sophi Antipolis, Frne {hnd Astrt Content-sed pulish/susrie systems offer onvenient strtion for dt produer nd onsumers, s most of the omplexity relted to ddressing nd routing is enpsulted within the network infrstruture. A mjor hllenge of ontent-sed networks is their ility to effiiently ope with hnges in onsumer memership. In our XNET XML ontent network, we hve ddressed this issue y designing novel lgorithms to speed up susription mngement t the routers, while gurnteeing perfet routing t ll times nd mintining ompt routing tles thnks to extensive usge of ggregtion. In this pper, we disuss the issue of susription mngement in ontent-sed networks, nd we speifilly desrie nd evlute the lgorithms tht we hve developed for XNET. 1 Introdution In ontent-sed pulish/susrie systems, messges re routed on the sis of their ontent nd the interests (susriptions) of the messge onsumers. This form of ommunition is well dpted to loosely-oupled distriuted systems with lrge onsumer popultions, with diverse interests, wide geogrphil dispersion, nd heterogeneous resoures (e.g., CPU, ndwidth). Severl tehniques hve een proposed to implement ontent routing, with vrious trde offs in terms of lgorithmi omplexity, runtime overhed, or ndwidth utiliztion. In prtiulr, support for perfet routing (i.e., messge trverses ommunition link only if there is some onsumer downstrem tht is interested in tht messge) introdues signifint mngement omplexity t the routers in the presene of dynmi susription registrtions nd nelltions. As onsequene, severl ontent-sed pulish/susrie networks do not systemtilly updte their routing tles upon onsumer deprture nd let the ury of routing degrde over time. In our XNET XML ontent network [7], we hve ddressed this issue y designing novel tehniques to speed up the most time-onsuming susription mngement opertion of the routers. Speifilly, we propose lgorithms tht llow routers to quikly determine the overing reltionships etween n inoming susription nd ll the entries of their routing tle. Covering reltionships re t the ore of susription ggregtion mehnisms, whih help limit the size of routing tles nd hene improve the effiieny of the filtering engine while ensuring perfet routing. Although the lgorithms presented in this pper hve een designed for XPth tree-strutured susriptions, they n e redily pplied to other susription lnguge with similr hrteristis. Experimentl evlution demonstrtes tht these lgorithms re highly effiient even when the numer of susriptions in the routing tle grows very lrge. 2 Relted Work Severl pulish/susrie systems implement some form of distriuted ontent sed routing, most notly IBM Gryphon [2], Sien [3], nd Jedi [10]. These systems dopt vrious pprohes to susription mngement. IBM Gryphon [2] uses distriuted filtering lgorithm sed on prllel serh trees mintined on eh of the rokers to effiiently determine where to route the messges. The system implements perfet routing nd supports susription registrtions nd nelltions; in ft, registering (neling) susription redues to inserting (removing) it from the serh tree nd is thus n effiient opertion. However, to mintin nd updte the prllel serh tree, eh roker must hve opy of ll the susriptions in the system. As onsequene, this pproh my not sle well to lrge nd highly dynmi onsumer popultions. Sien [3] lso uses network of event servers for ontent-sed event distriution, nd relies upon routing protool most similr to ours, ut with limited support for susription nelltion. In reent pper [4], the uthors of Sien introdue novel routing sheme for ontent-sed networking sed on omintion of rodst nd seletive routing. Susription mngement is simple nd effiient ut the system does not gurntee perfet routing, in the sense tht onsumers my reeive messges tht they re not interested in. The uthors hve ddressed this issue

2 y hving routers periodilly request for the routing tle of other routers. JEDI [10] proposes severl vritions for routing events mong its networked event servers; in prtiulr, with the hierrhil pproh, susriptions re propgted upwrd spnning tree nd messges re propgted oth upwrd nd downwrd to the hildren tht hve mthing susriptions. Susription mngement is simple nd effiient, ut this pproh my led to lrge routing tles t the root nd unneessry propgtion of events upwrd the tree. Our susription ontinment nd mthing tehniques re relted to the widely studied prolem of pttern nd regulr expression mthing. There exists severl indexing methods to speed up the serh of textul dt with regulr expressions, like the it-prllel implementtion of NFA [1] nd suffix trees [14]. In [6], the uthors hve ddressed the reverse indexing prolem of retrieving ll the regulr expressions tht mth given string. They propose RE-Tree, n index struture to quikly determine the regulr expressions tht mth given input string, y fousing the serh on only smll frtion of the expressions in the dtse. In [12], Tozw nd Hgiy present ontinment heking tehnique for XML shems, whih is sed on inry deision digrms. Little work hs een done on the prolem of ontinment heking for tree-strutured XPth expressions. In ft, the prolem hs een shown to e onp-omplete [11]. A sound ut non-omplete lgorithm hs een proposed in [5] to determine whether given treestrutured susription overs nother susription, ut it does not ddress the prolem of overing reltionships etween lrge sets of susriptions. 3 System Overview This setion gives n overview of the XNET ontent routing network. We lso riefly desrie its most essentil mehnisms, whih re relevnt for the rest of the pper. More detils n e found in [7]. System Model nd Definitions. XNET is distriuted pulish/susrie system whih onsists of olletion of ontent-sed routers (or nodes) orgnized in n overly network. Eh node routes the messges sed on its lol knowledge of the onsumer susriptions nd the tul ontent of the messges. Eh dt onsumer nd produer is onneted to some node in the network; we ll suh nodes onsumer nd produer nodes. We ssume tht ll routers know their neighors, s well s the est pths tht led to eh produer. We lso ssume tht the numer nd lotion of the produer nodes is known. From the point of view of router, this mounts to knowing whih neighors led to some produer. The onsumer popultion n e highly dynmi nd does not need to e known priori. The most reent implementtion of our routing protool, XROUTE, hndles multiple produers [8]; for the ske of simpliity, however, we only onsider networks with single produer in this pper. Eh node hs set of links, or interfes, tht onnets the node to its diret neighors. We ssume tht there exists extly one interfe per neighor, nd tht ommunition etween two nodes is relile. Our system lso inorportes fult-tolernt mehnisms to hndle oth trnsient nd permnent filures. <Quotes> <Stok> <Symol>DEF</Symol> <Prie>34.1</Prie> </Stok> <Stok> <Symol>GHI</Symol> <Prie>11.5</Prie> </Stok> </Quotes> () Symol // Stok Prie ="GHI" >"15" () Figure 1: () A smple XML doument desriing two stok quotes. () Tree representtion of smple XPth susription (//Stok[Symol="GHI"][Prie>15]) tht does not mth the XML doument. XNET ws designed to del with XML dt, the de fto interhnge lnguge on the Internet. Produers n thus define ustom dt types nd generte ritrry semistrutured events, s long s they re well-formed XML douments. Consumer interests re expressed using susription lnguge. Susriptions llow to speify predites on the set of vlid events for given onsumer. XNET uses signifint suset of the stndrd XPth lnguge to speify omplex, tree-strutured susriptions [13]. An XPth expression ontins one or more lotion steps, seprted y slshes (/). In its most si form, lotion step designtes n element nme followed y zero or more predites speified etween rkets. Predites generlly speify onstrints on the presene of struturl elements, or on the vlues of XML douments using si omprison opertors. XPth lso llows the use of wildrd (*) nd nestor/desendnt (//) opertors, whih respetively mth extly one nd n ritrrily long sequene of element nmes. We sy tht n XML doument mthes n XPth expression when the evlution of the expression on the doument yields non-null ojet. Figure 1 shows n XML event nd n XPth susription tht does not mth the event (eh rnh of the susription hs mthing node in the XML doument, ut the onjuntive ondition t the Stok node is not met). We sy tht susription S 1 overs or ontins nother susription S 2, denoted y S 1 S 2, iff ny event mthing S 2 lso mthes S 1, i.e., mthes(s 2 ) mthes(s 1 ). The overing reltionship defines prtil order on the set of ll susriptions.

3 The Routing Protool. XNET implements perfet routing, tht is, messge trverses ommunition link only if there is some onsumer downstrem tht is interested in tht messge. To do so, eh node in the network mintins in its routing tle olletion of susriptions tht desrie the lsses of messge tht its neighoring nodes re interested in. When reeiving messge, node first determines whih susriptions of its routing tle mth the event; it then forwrds the messge to ll neighors tht hve registered one of these susriptions. Given urte routing tles, this proess ensures tht messge eventully rehes ll the onsumers, nd only those, tht re interested in tht messge. When onsumer registers or nels susription, the nodes of the overly updte their routing tle ordingly y exhnging some piees of informtion tht represent the registrtion or nelltion of the onsumer. The proess strts t the onsumer node nd termintes t the produer node(s), following the shortest pths. As onsequene, messges pulished y the produers follow the reverse pths of the susriptions, long multist tree spnning ll interested onsumers. The routers in our system redue the size of their routing tles s muh s possile y using elorte ggregtion tehniques, whih re sed on the detetion nd the elimintion of susription redundnies. Susription ggregtion llows us to drmtilly improve the routing effiieny of the system oth in terms of throughput nd lteny, euse the time neessry to filter messge is proportionl to the numer of entries in the routing tles. On the other hnd, ggregtion lso dds signifint omplexity nd overhed to the routers, euse they need to identify the overing reltionships etween inoming susriptions nd ll the entries of their routing tles. These mngement opertions were the min ottlenek of erly XNET implementtions nd led us to develop the tehniques presented herefter. More detils out susription ggregtion nd XNET s routing protool n e found in [7]. 4 Susription Mngement Effiient susription mngement is ritil for the overll performne of the system nd to gurntee short registrtion delys to onsumers. As previously mentioned, the ost of susription mngement minly results from the extensive overing heks tht hve to e performed y the routers when susription is registered or neled. To determine whether given tree-strutured susription lso lled tree pttern heneforth overs nother susription, we n use the lgorithm proposed in [5], whih hs time omplexity of O( S 1 S 2 ), where S 1 nd S 2 re the numer of nodes of the two susriptions eing ompred. 1 Oviously, when n inoming 1 Note tht the overing prolem hs een shown to e onpsusription must e tested for overing ginst ll the other susriptions in the routing tle, itertive exeution of the lgorithm is lerly ineffiient. We hve therefore designed novel lgorithm, termed XSEARCH, whih effiiently identifies ll the possile overing reltionships etween given susription nd possily lrge set of susriptions. This lgorithm is desried in the rest of this setion. Additionl detils nd proofs n e found in ompnion tehnil report [8]. Prolem Sttement. Consider tree pttern s nd set R of n tree ptterns, R = {s 1,, s n }, whih we will refer to s the serh set. Our lgorithm runs in two different modes ording to the reltionships tht we wnt to identify. Covered mode identifies the set R of ll the tree ptterns in R tht re overed y s. Cover mode identifies the set R of ll the tree ptterns in R tht over s. We refer to XSEARCH nd XSEARCH s the lgorithm running in overed nd over mode, respetively. Definitions nd Nottions. Let u e node of tree pttern s; we denote y lel(u) the lel of tht node nd y hild(u) the set of the hild nodes of u in s. Rell tht the lel of node u n either e wildrd (*), n nestor/desendnt opertor (//), or tg nme. We define prtil ordering on node lels suh tht if x nd x re tg nmes, then (1) x // nd (2) x x iff x = x. Algorithm 1 dd(s, t, u) 1: if t hild(t) suh tht lel(t ) = lel(u) nd s / su(t ) then 2: su(t ) = su(t ) s 3: else 4: rete t hild(t) suh tht lel(t ) = lel(u) nd su(t ) = {s} 5: end if 6: for ll u hild(u) do 7: dd(s, t, u ) 8: end for Ftoriztion Trees. Our XSEARCH lgorithm does not operte diretly on the set of tree ptterns R, ut on ftoriztion tree uilt from the set R nd defined s follows. The ftoriztion tree of R, denoted T (R), is tree where eh node t hs two ttriutes: lel lel(t) similr to tht of node of tree pttern, nd set of tree ptterns su(t), whih is suset of R. The root node r T of T (R) hs no lel nd su(r T ) = R. Initilly, T (R) onsists of only its root node r T. We inrementlly dd eh tree pttern s R to T (R) with the reursive dd(s, r T, r s ) funtion shown in Algorithm 1, where r s is the root node of tree pttern s. The removl of tree pttern from T (R) is performed in similr mnner using Algorithm 2. Note tht, to keep the presenttion simple, we omitted the speil omplete [11]. Our lgorithm is sound ut not omplete, i.e., it my fil to detet some overing reltionships in rre pthologil ses, ut ll the reltionships tht it reports re orret. Consequently, router my fil to ggregte some vlid susriptions, ut orretness is never violted.

4 se of the root node of the ftoriztion tree in the ddition nd removl lgorithms. Algorithm 2 remove(s, t) 1: for ll t hild(t) suh tht s su(t ) do 2: su(t ) = su(t ) \ {s} 3: remove(s, t ) 4: if su(t ) = {} then 5: remove t from hild(t) 6: end if 7: end for R = s 1 s 2 s 3 s 4 s 5 s 6 * {s 5 } // {s 1,s 3,s 4,s 6,s 7 } 3 * {s 2 } 1 11 {s 5 } 2 {s 1 } {s 3,s 4,s 6,s 7 } {s 4 } {s 2 } r T {s 3,s 4 } {s 3,s 6 } {s 6 } {s 4 } // {s 1,s 2,s 3,s 4,s 5,s 6,s 7 } Figure 2: Six tree ptterns nd orresponding ftoriztion tree, where node is represented y its lel. Eh node is ssoited with set of tree ptterns, shown etween rkets. Intuitively, ftoriztion tree enles us to remove the redundnies etween the tree ptterns in R y ftorizing identil rnhes. Thus, T (R) is ompt representtion of the tree ptterns in R. Figure 2 shows n exmple with six tree ptterns nd the orresponding ftoriztion tree. It is importnt to note tht the ftoriztion tree is not unique; depending on the insertion order of the tree ptterns, we n hve distint, equivlent trees. This does not ffet the orretness of our XSEARCH lgorithm, nor its performne. s 7 The Serh Algorithm. We first desrie the XSEARCH lgorithm in overed mode. Consider susription set R nd orresponding ftoriztion tree, T (R). Let s e single tree pttern. The lgorithm works reursively on the nodes of s. When exeuted with the root nodes of T (R) nd s, XSEARCH (r T, r s ) returns the set R of ll tree ptterns tht re overed y s. The serh proess is desried in pseudo-ode in Algorithm 3. Intuitively, it tries to lote the pths in T (R) tht re overed y s; the tree ptterns tht the union of those pths represent re lso overed y s (lines 6 nd 8). The proess is slightly more omplex when enountering n nestor/desendnt opertor (//), euse we need to try to mp it to pths of length 0 (line 11) or 1 (line 12). To etter illustrte the workings of the lgorithm, onsider the exmple runs shown in Figure 3. Two tree ptterns, u nd v, re mthed ginst the ftoriztion tree T (R) of Figure 2 for lrity. The nodes of u, v, nd T (R) re numered in the figures; we refer to node numer i of u, v, nd T (R) y u i, v i, nd t i, respetively. The different steps of the lgorithm re detiled in the two exeution tres (the symol represents reursive invotions of the lgorithm). Algorithm 3 XSEARCH (t, u) 1: if t is lef then 2: XSEARCH (t, u) = 3: else 4: if lel(u) // then 5: if u is lef then 6: XSEARCH (t, u) = t hild(t) lel(t ) lel(u) su(t ) 7: else 8: XSEARCH (t, u) = t hild(t) lel(t ) lel(u) 9: end if 10: else 11: S 0 = u hild(u) XSEARCH(t, u ) 12: S 1 = t hild(t) XSEARCH(t, u) 13: XSEARCH (t, u) = S 0 S 1 14: end if 15: end if Algorithm 4 XSEARCH (t, u) 1: if u is lef then 2: XSEARCH (t, u) = su(t) 3: else 4: if lel(t) // then 5: if u hild(u), lel(u ) lel(t) then 6: XSEARCH (t, u) = su(t) 7: else 8: if t is lef then 9: XSEARCH (t, u) = 10: else 11: XSEARCH (t, u) = 12: u hild(u) lel(u ) lel(t) 13: end if 14: end if 15: else 16: S 0 = t hild(t) XSEARCH(t, u) 17: S 1 = u hild(u) XSEARCH(t, u ) 18: XSEARCH (t, u) = S 0 S 1 19: end if 20: end if Algorithm 5 XSEARCH (r T, r s ) t hild(t) XSEARCH (t, u ) u hild(u) XSEARCH (t, u ) 1: XSEARCH (r T, r s) = su(t) \ t hild(t) XSEARCH (t, r s) The seond lgorithm, XSEARCH, is desried in Algorithms 4 nd 5 nd works in very similr mnner. The mjor differene is tht the lgorithm works reursively on the nodes of T (R), trying to find pths in s tht re overed y the tree ptterns in T (R). The reursive funtion in Algorithm 4 returns the susriptions tht do not over s. A susription t overs s if eh rnh of s is overed y some rnh of t (line 12). Susriptions tht hve longer (line 2) or inomptile (line 6) pths nnot over s, wheres shorter pths (line 9) re eptle. Finlly, when enountering n nestor/desendnt opertor (//), we need to try to mp it to pths of length 0 (line 16) or

5 * u v // XSEARCH(r T, u 1) = XSEARCH(t 3, u 2) XSEARCH(t 3, u 4) XSEARCH(t 3, u 2) = XSEARCH(t 4, u 3) XSEARCH(t 5, u 3) XSEARCH(t 9, u 3) XSEARCH(t 4, u 3) = XSEARCH(t 5, u 3) = {s 3, s 6} XSEARCH(t 9, u 3) = {s 4} XSEARCH(t 3, u 2) = {s 3, s 6} {s 4} = {s 3, s 4, s 6} XSEARCH(t 3, u 4) = XSEARCH(t 5, u 5) XSEARCH(t 9, u 5) XSEARCH(t 5, u 5) = {s 6} XSEARCH(t 9, u 5) = XSEARCH(t 3, u 4) = {s 6} Finlly: XSEARCH(r T, u 1) = {s 3, s 4, s 6} {s 6} = {s 6} XSEARCH(r T, v 1) = XSEARCH(t 3, v 2) XSEARCH(t 3, v 2) = S 0 S 1 S 0 = XSEARCH(t 3, v 3) S 1 = XSEARCH(t 4, v 2) XSEARCH(t 5, v 2) XSEARCH(t 9, v 2) XSEARCH(t 3, v 3) = su(t 5) su(t 9) = {s 3, s 4, s 6, s 7} XSEARCH(t 4, v 2) = XSEARCH(t 5, v 2) = S 0 S 1 S 0 = XSEARCH(t5, v3) S 1 = XSEARCH(t6, v2) XSEARCH(t7, v2) XSEARCH(t8, v2) XSEARCH(t 5, v 3) = su(t 6) = {s 6} XSEARCH(t 6, v 2) = XSEARCH(t 7, v 2) = XSEARCH(t 8, v 2) = S 0 = {s6} S 1 = XSEARCH(t 5, v 2) = {s 6} XSEARCH(t 9, v 2) = S 0 S 1 S 0 = XSEARCH(t9, v3) S 1 = XSEARCH(t10, v2) XSEARCH(t 9, v 3) = XSEARCH(t 10, v 2) = S 0 = S 1 = XSEARCH(t 9, v 2) = S 0 = {s 3, s 4, s 6, s 7} S 1 = {s 6} XSEARCH(t 3, v 2) = {s 3, s 4, s 6, s 7} {s 6} = {s 3, s 4, s 6, s 7} Finlly: XSEARCH(r T, v 1) = {s 3, s 4, s 6, s 7} Figure 3: Two XSEARCH exmple runs. 1 (line 17). Note tht we impliitely introdue n rtifiil root node in the tree-strutured susriptions (denoted r s for susription s) in order to simplify the desription of the lgorithm. When lled with the roots of the ftoriztion tree nd susription s, Algorithm 5 reusively serhes for susriptions tht do not over s nd return the omplement set with respet to R. Beuse of spe limittions, orretness proofs re not given here (they n e found in [8]). Both Algorithms 3 nd 4 perform in O( T (R) s ) time, where T (R) is the numer of nodes in the ftoriztion tree nd s tht in the expression eing tested. This qudrti time omplexity is due to the ft tht eh node in T (R) nd s is heked t most one. As for the spe omplexity, the size of the ftoriztion tree T (R) grows linerly with the numer of tree ptterns in the serh set R. However, y onstrution, the ftoriztion tree typilly requires muh less spe thn would e needed to mintin the whole serh set R, tht is, T (R) s i R s i when R grows to lrge vlues. 5 Performne evlution Experimentl Setup. To test the effetiveness of our susription mngement tehniques, we hve onduted simultions using rel-life doument types nd lrge numers of susriptions. We hve generted relisti susription worklods using ustom XPth genertor tht tkes Doument Type Desriptor (DTD) s input nd retes set of vlid XPth expressions sed on set of prmeters tht ontrol: the mximum height h of the tree ptterns; the proilities p nd p // of hving wildrd (*) nd nestor/desendnt (//) opertors t node of tree pttern; the proility p λ of hving more thn one hild t given node; nd the skew θ of the Zipf distriution used for seleting element tg nmes. For our experiments, we generted sets of tree ptterns of vrious sizes, with h = 10, p = 0.1, p // = 0.05, p λ = 0.1, nd θ = 1. We used the widely-used NITF (News Industry Text Formt) DTD [9] s the input DTD of our XPth genertor. All the lgorithms were implemented in C++ nd ompiled using GNU C++ version Experiments were onduted on 1.5 GHz Intel Pentium IV mhines with 512 MB of min memory running Linux XSEARCH Effiieny. We evluted the effiieny of the XSEARCH lgorithm for serh sets of different sizes. For this experiment, we onsidered serh sets with unique susriptions, tht is, given susription does not pper more thn one in set. Indeed, in given router, XSEARCH is used to determine the overing reltionships etween given susription nd the susriptions in the routing tle, whih re ll unique. For eh serh set, we generted 1, 000 dditionl susriptions nd, for eh of them, we mesured the time neessry to determine the suset of the susriptions tht over, nd re overed y, tht susription. For omprison purposes, we hve lso mesured the effiieny of the XSEARCH lgorithm ginst sequentil exeution of the ontinment lgorithm of [5], whih we ll Liner. Figure 4 shows the verge serh time of the XSEARCH lgorithm. It ppers lerly tht XSerh is extremely effiient. Even for very lrge serh sets, we n expet n

6 Time in ms Averge serh time for XSerh Averge serh time for XSerh Size of serh set Figure 4: Averge serh time for the XSEARCH lgorithm. verge serh time of less thn 50 ms. In omprison, the Liner lgorithm yields to serh times tht re systemtilly more thn two orders of mgnitude higher. This result is not surprising, s the Liner lgorithm needs to evlute the entire susription set R while XSerh only serhes through the ftoriztion tree, whih is muh smller y onstrution. The seond vrint of the lgorithm, XSerh, tends to e slightly less effiient thn XSerh for lrge onsumer popultions. We n explin this oservtion y the ft tht, on verge, the XSerh lgorithm neessittes deeper trversls of the ftoriztion tree. Size of serh set 1,000 2,000 5,000 10,000 XSEARCH XSEARCH Tle 1: Averge serh time of XSEARCH in ms. One should note tht, in prtie, the sizes of the routing tles rrely exeed 1, 000 entries, even for very lrge onsumer popultions, thnks to susriptions ggregtion. For ompleteness, we show in Tle 1 the solute verge serh time of XSEARCH for serh sets of smll sizes, whih re most relevnt in the ontext of ontent-sed routing. R 1,000 2,000 5,000 10,000 20,000 50, ,000 s i R si T (R) Tle 2: Spe requirements for given susription popultion R nd its ftoriztion tree T (R), in thousnds of nodes. Spe Effiieny. We hve experimentlly quntified the spe requirements of the ftoriztion tree with susription sets of vrious sizes. The results in Tle 2 onfirm tht the numer of nodes in the ftoriztion tree is indeed notly smller thn the sum of the nodes of the individul susriptions. 6 Conlusion We hve desried the susription mngement tehniques tht we implemented in our XNET ontent routing network. These tehniques rely on XSEARCH, n lgorithm tht determines the overing reltionships etween susriptions, to effiiently proess onsumer registrtions nd nelltions. By pitlizing the performne of this lgorithm, our ontent-sed pulish/susrie system n mintin ompt routing tles (for improved routing performne) while ensuring perfet routing (for ndwidth effiieny) t ll time. Although desried in the ontext of ontent-sed routing nd XPth, the XSEARCH lgorithm n e redily used with similr susription lnguges or to ddress different dt mngement prolems. Referenes [1] R. A. Bez-Ytes nd G. H. Gonnet. Fst text serhing for regulr expressions or utomton serhing on tries. Journl of the ACM, 43(6): , [2] G. Bnvr, T. Chndr, B. Mukherjee, J. Ngrjro, R. Strom, nd D. Sturmn. An effiient multist protool for ontent-sed pulish-susrie systems. In Proeedings of ICDCS, My [3] A. Crznig, D. Rosenlum, nd A. Wolf. Design nd Evlution of Wide-Are Event Notifition Servie. ACM Trnstions on Computer Systems, 19(3): , [4] A. Crznig, M. J. Rutherford, nd A. L. Wolf. A routing sheme for ontent-sed networking. In Proeedings of IEEE INFOCOM, Mr [5] C.-Y. Chn, W. Fn, P. Feler, M. Groflkis, nd R. Rstogi. Tree Pttern Aggregtion for Slle XML Dt Dissemintion. In Proeedings of VLDB, Aug [6] C.-Y. Chn, M. Groflkis, nd R. Rstogi. Re-tree: n effiient index struture for regulr expressions. The VLDB Journl, 12(2): , [7] R. Chnd nd P. Feler. A slle protool for ontentsed routing in overly networks. In Proeedings of NCA, Cmridge, MA, Apr [8] R. Chnd nd P. Feler. XNet: An XML Content Routing Network. Tehnil report, Institut EURECOM, [9] I. P. T. Counil. News Industry Text Formt. [10] G. Cugol, E. D. Nitto, nd A. Fugett. The JEDI eventsed infrstruture nd its pplition to the development of the opss wfms. IEEE Trnstions on Softwre Engineering, 27(9): , Sept [11] G. Miklu nd D. Suiu. Continment nd equivlene for n xpth frgment. In Proeedings of PODS, Mdison, WI, June [12] A. Tozw nd M. Hgiy. XML shem ontinment heking sed on semi-impliit tehniques. In Proeedings of the Conferene on Implementtion nd Applition of Automt (CIAA), July [13] W3C. XML Pth Lnguge (XPth) 1.0, Nov [14] S. Wu nd U. Mner. Fst text serhing: llowing errors. Communitions of the ACM, 35(10):83 91, 1992.