This paper proposes a. iiietliotlology for perforniing an evaluatioii and optimization of tlip cost of an ,4TM swit.cliiiig architect,ure under perfornia.nce constraints given in t,eriiis of virtual colinect.ion lilocking probability.An a.nalysis of blocking networks is de~-eloped, a.nd combined wit,li known results concerning n o w blocking networks. provides a t.heoret,ical iiiodel which rehtes traffic characteristics, network topology and blocking prolia1ilit.y in a mult.iiat.e/iiiul-Biservice l~roadlmiicl environment.. A n a.nalysis of the characteristics deteriiiining t,he cost of a generic AT11 swit,cli iiiipleiiiciitatioii follows. The niotlel is oriented to optiniize Iiotli t,lie topological para.iiiet,ers and tlie s p e d atll-ailt,a.ge. with respect t o t,he main cost factors of l'LSI-based swit,cliing net\vorlis i.e. coniponmt s count and coniplexi t y. iiibercoiinect ic ni costs.
Research scopeScope of t,he research is a particular class of AT11 sTl-itches. i.e. those with large mnlt ist age huffered int,erconiiect ion This pqier is focused on tlie anal) 10148 Torino, Italy networks, providing multiple paths Iiet,meen a.ny iiilet /out.let pair. In the paper we point 0111-a,tteiitioii towards Clos networks. but. tlie methodology is general, anti ca.n lie iniinedia,t,ely qqilied to other multistage topologies. The practical iniport,a.nce of tliis family of networks in the AThI field is deiiioiistrat.ed by their use in most. of tlie current. swit,ch prot ot.ypes [1.10,15.28,12,2O].Mult,istage iiet.morlcs are manda.tory even for sniall size switches, due t,o t.ec1inology limit,ations, topological considerat,ioiis and cost,. In soiiie ca.ses niultipatli fabrics e nploy internal data.gra.m routing, i.e. ea.ch cell is indepeiidently rout.ed to its destination, while in other solut.ioiis all cells of a particular connection follow the same pa,t,li ("connection orient,ed" fabrics): in these cases virt,ual colinecbion blocking is possible.Although t,lie theory of call blocking iii circuit switcliing t,eleplione net,works is well est~ahlislietl, these results ca.n not, be a.ppliec1 directly to deriw virtual circuit ATM hloclting probability, a.s tlie inult,ira t,e environment introduces new variables t,liat. lead 6o different results.we do not take int,o a.ccoiint prolilenis related to cell loss prolialiilit~y. liantlwitlt~li allocat.ion for va,riable bit rat,e connections and real time back pressure congestion control. In the following sections w e will asslime that, ea.ch coiiiiectioii requires a cert,aiii a.mount, of bandwidth. either being tlie actual lmiclwit,li of a coiistaiit bit, rate (CBR) connection. or tlie "effective" liand~vidt~li of a varialde bit rat,e (VBR) connection deteriiiiiied by some resource alloca.tion algorit,lim.Congestion control is realized by tlie resource a.llocation policy, assigning a cert a.in haiitlwit,li t o l'BR connect.ic.nis: policing devices locat,etl a.t the user interface are responsible for enforcing tliis resoiirce allocat,ion policy. while buffers a,nd back pressure ni...