Tyand suppose T; starts execution at Ame ft Toeks Sy Asn 157i intated_ and it peeps T, ong to is higher priority. At time 1,7, locks $- At me 1,7 altempl 19 lock S, bu is blocked because, has finished with it Ty 5 which now inherits the prioity of Ty starts executing ioweve, when atime res o lock Sit eanot do so since has Tock nit Both, and T, are now deadlocked. “There is anaher drawback of priority inheritance. It is possible forthe high Fniorty tsk to be blocked once by every ober tsk executing, on the sam processor. ‘To get around both problems, we define the priority eiling protocol. T| plot cling ofa semaphore i the highest priority of any task that may look Lat PT) denote the prioity of task T and P(S) the priority elling ofthe seman oferta eatin, = late sen Baer ee eet le ei ey eng OUD 8 ae we ay ask 638 al80 be blocked frm eeqng et i corre held by sme other at ya Ca i, hone 1 the priority of T ony cigs gra y ropa Fh ry in a sing, “re pity cling protocol prevents desde et B, be the st ofall ei sections hal can ease te ‘and 1(2 be the time ten for seston to beeen ene mx ted. The, Tw tHocked frat most" ), ‘The highest pron the processor (Le. it is blocked) whene a er seeks 10 ack the semaphoisgurdng «citi secon whichis aeay ey some ther tke Q (in hich case itis nid we ae ‘or when there exists a semaphore $! Hosked by tsk), sk Tsai POcessr.Telngies locked by ory exling eon whose priority ecling or gett than or ql the pty off er uote aera pore tow oaety weer ta oon re i tly ua cmay sua er ‘Suppose. blocks one or mare sks. Then it aber the pity task that it" is cumently. blocking, The operons of pity inheritance and resumption of previous pray rind, Priori ineritanee is transitive, ‘A task Ty can preempt another tsk T, iT, oe ot ol el seotion which T, curently needs and ithe cures ity of rater than that ofthe curetpioity oT Fig, 4.10, Priori cling preaem alin FeSBTD ements wing He rin ling pes 6:2.2. Pamemtn Eatin Denti FZ IE Bigot fs gcsne slong, See TOF sprites io oninte tative is ie extice. EDF in 2 Grate perky Sebo wipers, so Gna spe Sep, ee ech piecing der ‘Soe ade UD) ie lit Se aie et ‘exargie ‘Comite te folowing vt of (asin eave to ween. [Teme | “Fork | Arrival Vine | Execution Time | Absolute Deadline | o w 3 2 3 10 7 fs w as] Tin oly ak wing ran wn vo as er latipes pty an 1 wna Ty ave immcditely Y aives atime 4 vce tT, aives a tne 3; however since dy > a, slower privity Ua fr Fk finish. When T iiss (at time 7). T start (sine at which point ean resume aot a) 11 and must has higher pony dT). run un 1 to completion & EDF ts an opti! wigrocessor Seheduling algorithm, That is, i EF cannot feasibly schedule task se on # uniprovessr, there is 19 ot setedangalgovitin that ean. & Ifa the tks are periodic and have relative deadline equal 1 te pets the test fr tak, Set schedulabity i particulary simple the total ilzation ofthe task set no greater than 1, de ask et ‘be easily schedule on a ingle processor by the EDF algoritun. Scheduling Test for EDF Define U= Ee, 7P.) daar ™ 2H, {dl} and P= LEM (ys Pa Zo wth increased service RIS) a, Mie reward funtion associated with a0 TS skier se given tit Typically the reward function isos fan ib the ote 0 ite 04m tere r(x) is monotonically nondecessog in, : “Te reward sO pt sometime mf tease : if es isnt cued pia point, it produces no useful output. 2s “asks with this evar finction cn be canbe end savings mad ‘and an optional component, | ‘The mandsory portion mus be completed by he deine i be kl eric te optional poon is dane tine permis ‘The epional porn egies ttl of tnes compl. nex se th exeetion ofa ask mse oped yi deed. Identical Linear Reward Functions Fertak ,:the reward function is given by 0 itxsm R(x) = 4 x= m itmers0.+m, 0, ifx>m +0,04s Oy possible 1 (38) Enbedied ond Real Fro the reward from excong anit of epina work i one mit Agere ssid to be optimal if the reward is maximized subject to all tasks complery their mandatory portions by the task deadline. 7 The tasks T, ... T, have mandatory portions My, bed ee Mand option poy ey = Define M = {Myo My) 0 = (0-0) Te TT) “The opal algorithm for IIS Ii shown in lowing Sau. We re unit of reward for-each unit of the optional portion coimpleted for any tak pict reward fs obtained when the proceor caries out a8 much exseuiay ‘Run the EDF algorithm on the tisk set T to generate a schedule, y ‘this is feasible, ‘An optimal schedule has been found: STOP. ise, go'to step 2. endif [Run the EDF algorithm on the task set M, t0 generate a schedule, this set snot feasible, “Teannot be feasible scheduled: STOP. : Elie, Define a, a8 the fh instant Sq when either the’schedled tsk changes, or the processor becomes le, 1,2 Let k be the total number ofthese instants Define ay a8 when the frst task begins executing i Sq Define x) asthe task that exeoutes in Syn 24 +3} Define L, () atid Ly () 2 the total exeeution — ‘a0 0 3. ‘edit a sek} “go while(0m, #05 Ech task has weight , astocited with it Astine tat he esha Inder in noninereasing order of weighs, 2,2. 2e Th ose -cnsly scheduling such tasks is obvious. ‘Alvays Rum the available tsk with the getest weight ue 1 th ed south mandatory portions ofl the tasks byte respi dein. TH forty algorithm ARIS 2, which is shown in followin fuimbedded ant Real Ti Mel 7 sions of all the asks, ang wera we wane a Se anders ete roaig che ts mt the task set Tis nt schedulable: STOP. che do while 54m se 0'= OU 10. Define 0, to be the prt of seuMi= Mu (0,1 and use IRISH to Find an opin chess, scheduled by IRI peed endo cendit cal As with IRIS, we check to see i ‘mandatory portions. Hat, we tp i & Ite sueeok, we proceed by Roiing II fsa tothe mandatory potions ofthe tasks and he Set of option omtions ua nly to optional portion of ask T if we can feasibly schedule alle 5 aay 1 ith a mandatory tok 45.22.0/1 Reward Functions We asume hoe that fran task / he reward function fs given by 0 ifs sm, +0, | ie) we get no rewan! for exciting the optional portion pataly. IF we ren optional potion to completion, we obsin one uit of reward; tennis We & pothing {fan optional pt is not runt competion i 5 tion in the eu sheds, we @ We continue in man the dealin “phe optimal SIRICEY WOU be 9 omy & TM ible, subject tthe constant ar ns jos must be met, » erin orton of al he mandatory Giforunately, when the eXecution tings Souining an optimal sedate can be sho as ai 2, the pectlem, nnd NPComune selena oy ‘The IRIS3 algorithm is based onthe following pen ‘Therefore, We assign eight acconting tthe in the optional portions and run IIS? "of he duration of] remove its optional portion fom consitaton and eum nn i mer unl each optional portion bss been ei scheduled to completion or dropped ahgether bee eer Run the EDF algorithm onthe et M of mandatory sks FM is not EDF sehedulal Task setT cannot be feasibly scheduled STOP. ele Gow sep 2, endif iis the st of optional ponions Assign, = 1/0, fori = 1.0m Renumber the tasks so that thir weights se in 8 non-sscenting sequence, ie. wp 2w ey Run algorithm IRIS? on a task set composed of the mandstory set M and optional et Oto bin schedule Spaa Stems ia te previous SECON, We consider a Eablet Ral Tne 5a Fall the optional asks in O are executed to competion in Sp in ks Return Sy and STOP. 1s i tor portions of 20. There a mae re eines an ke pa pe owed intone en by ‘ll. We sume that fray ‘Let fy be the smallest index # such that = Res {@ itosx so, 7 (0s not rum to completion in Sy : 4,10) itzz0,” fa “ he functions f, ae one-one and concave, eno the service that as T, restive "The total amount of service that us Redefine O=O~ (Org) 1 0) Letty the inte |accule S LT ne, D)antr ‘ optimization ogi alee § xj 8. The onininaton problem tecorekaertemerinag nizing Gorosep3. endif ead Fig 412 Algor TRIS ‘ P= Z 4 45.24 Identical Concave Reward Functions rc) 1 tothe constrains that In this case, we consider tasks with identical release times and with mandsion bis Ex = DD, orton fo, We assume tate ead funtion oF, is given by fare Ro = fl, Osx=0, ey ee 0) 1) ~ [yoy ifx= 1 ay 20 tsjsise ce oes eae ee ling Lagrange multiplier. The‘soltion 10 tit pinieaon pace coe oa ry 8 bin po | multipliers. ¥ ° were the functiof is one-to-one and concave . Recall that a function /() oneave iff forall, and 050 flan +[-a}xy) 2 afte) +See) CGeomettialy, this condition can be expressed by saying that, for any to points on concave curve, the stright line joining them must never be above f(D) ~My Isjstsn ol ere. An example of concave functions 1% : ana fenaina | Exy-O-d.y = 0 on : “We will also assume that the functions f(x) are differentiable, and define ms . 43) ete) =a) ds. We wl asume tha theives uncon gf g exis era 320, vy4y=0 ¥, 30 lgystsn «| ere from above equation, 50 and so fom equation () 81660) 5 lajsisn ty) «| 5120, we have ny = 0 and so fom equation) L £6) = 4 Isjsisn C) ‘hie wil happen f the fnctions gare monotonically decreasing and we asun| that they ae. The tasks are numbered in néndeereasing order of thet abso deadlines (1: D, SD, $... sD.) For notational convenience, define Dy 0Eanbeed ond Rel ine Sp (ed) Let (and 3, ‘enotes values for 50) andy (6 = Vm that sai equations () 10 (6). 4.5.3. TASK ASSIGNMENT ites typially allocate according to some simple eiteion and hope thy that criterion, When checking a alloca ei feasibility wil follow asa side fect of @ easily, we must account for communication cos For example, suppose tat T, < Ty Task T; cannot stat before receiving tek, opt Tati, denotes the completion tine of sk T, and, se time o communist fom , 07, rR hiten tasks T, and T) at allocated to the same pressor, then ¢1» = 0. I they ae nis postive and must Be taken no aso for located to separate processors, hile eeeking for feasibility 4.5.3.2. Utilzation- Balancing Algorithm ‘This algvitim ottempts to balance processor uiization, and proceed by locating the tasks one by one and selecting the least utilized processor. Te gorda is shown a following figure. “This algo tes into account the possibility that we might wish 1 ra mukiple copies of same task simultaneously for fault-tolerance. Tn particu, i fssina r,eopies of tsk T, to separate processor. Let uj and uP denote th izations of processor p, under an optimal elgoridim that minimizes the sum of the squares of the processor utlizations and under the bestfirst algorithm, respectively. Ir, =. ra = rand there are p processors inal itis possible stot Lory yr BON no sig fap (rr itp >> this rato ends to 1.125, whichis agreeably smal eters foreach ask Te do Allocate OOP OF ek ho update the Processor allocation to acesun o a fp ‘tized rocesons easton fi, end do (ober 3s the redundancy be scheduled). the mun HE Of copies of task, that mua Fig, 413. Uitetionbtercing gery Nextt Algorithm for RM Scheduling sao “per isa utlization based llcation bei hat is mean oniuaction with the remem tin conjunction oie sheng cacti assed 0 consis fen procera resources ote han frei mo oss ine: Defite Mo sci placed by thes Tak Tsincasy engin 21s emt secialy be loom, The eS |at sony allocated the tasks ofthat class, | | We allocate tasks one by one to the app ce en ot |schedulability. om esis sete ie | eee we have a set of periodic independent pretmplible asks to be esiped| 1 alroceno consisting of en pce, Tet eis al creo ter tan proceso tines guarana Tad ol eal Fa tions of the tasks =] “We mow tn ong a sum of te tions oF 1 Sk igng processor so greater than 1, the tsk sets EDF schedulable on that proego eee oblemstucs to making tk sent Wh the pope thu sriat tendon of the sk signed oa procsor doesnot exced |.) ‘Wert ke minimize he ier of processor needed, This sinpaking problem and many slots exist fer solving “Te alot we present here te ist dereasing ago, Sup eat tobe snpned. Prepare sorted Tit L ofthe tasks so tas zane noinereasing ore. This pode o sow tat when he ber of proctor FU i ge,y he fama "Number of pocsessos sed by the fist ft decreasing algorithm “Namber of processors used by optimal algoriim approaches 11/9 ~ 1.22, when 2 large task set is used. In fac, this lini approwched quickly, so that 1.22 is a good measure even for relatively sat ome 15:34. AMyepcoffn Sehedllng MS) Agr ‘Tus fe fe toned ht ss canbe preempted The mo Dine hcg hun s on esinnentedling sein my fen peenpve was Moni offing steung get thes acount no cay of proceing nets ut loo any eget th ak yh {Sakon resources > Por itn, atk may ned be exci ees 0 kl amr rma eda hve ntl oe prt. © MOS wan ofine grin nat we re given in avant x fuk iran tines, execlon tines and edie MOS proceed by bling asco tes ‘Bach node ins re reprsens an iinet and shed ube a he ks, ° seems ee © Bach child of a noag extended by on tn 4 Aleafof this te consi sis ote «pe stile 8 TO AD tt yy an aa ont of 04 PSs chet! Ley se pastincan be esi 5 ls ty teas We tart a the 001 node, which ink having Ben sche one hes ton0 tak having Been sede ie) tea 2) Pree 0 buld he te foe ht poi anode ns developed a flows, (0 ahe rot oft eau —_ 3 Teron Shera ; tt ene age 1 deep cade, Given a node n, we try to extend by one more task 2, We pick up one of the as yet us schedule represented by node n, 3, The augumented schedule is acid aod fn ‘There are two questions that mst be answered, U8 Sede represen by ta nd edule tasks andy to a it he 1) Which tsk do we pik for extending an incomplete scheduler Ans: The Wk tht We ce eet " fa esp se ce shinies te eo : saya fn Task execution time ens © Destine aris stor time $ Laxiyor icihted sum ofany of the above For instance, if H(() = D, , then the next tsk tobe chosen for scheduling will be| HHeaeyetcunscheduled task with the carliest deadline.Embedded and Real Ting ae 2. When do we decide that » node is not worth developing further yay to another node? ny Ans FE nw ma hp asl and develo i Taga "fgg ap —— fat icc Moras 2a Cia PK 15100 FAFEE, We Wl spe g 4 We nly develop a node iit is strony feasible 2, feasibility of modes. Ft oon 4 A node is strongly easibe if feasible sthedle C4 be Lene . eee | festonding-the curent pail schedule with any one of the ans? j.Foeed AUUFSHNE and Bide AB Serre et aged addressing and bidding (rap) ye dure and 5 sed fr tsk sts conser nn cr sing te 8H oe an Ona eaay mien ten paees eae ‘that are leaves can represent a feasible schedule. "s ttweesenners ne tnt en fase enki ay | SP eauinesiap eet eeaS aera cn aeny wif ei facets en se aieeeg airman reer |e will need to check feasibility oF extending the schedule by each of te ‘asyet- unscheduled tasks, As a result, the number of comparisons need los fo FAB elgorithm i sed when proceso, deine jeresares oF He 1 EXECU the tk In that cae nae [cone rroeworin the system. The FAB ler nec Sp Bich oer minis taste ar ine ee ee the mumber of tasks is a is ve evel generate one root-to-lat path is ele y+ oe “These include the Set of cris tasks and Any aon that it may have accepted, 1 mainains a able ofthe spl compan 5 oa apc to processor in the system, a al net yD # me | ‘To reduce the number of comparisons, we ean replace the strong fexsibiy check al cach node by means of 8 myopie procedure as follows. is procedure picks the frt mi ‘The time axis is divided into windows, which ae itenas of fice ‘uration and each processor regulry sens to its collages the aston ofthe next window that is curently fe Since the system is distributed, tis information may acne be completely up to date 8 shopping for a processor on which to fled a task, an ove checks its surplus information and sles a processor ale the Foessed 1 For ach nom leaf level ~ node. {Kony} se-yetsnscheduled tasks 2. Checks o se ithe schedule represented by m can be feasibly exten by each ofthese tasks. act ws mark the ode as hopless and backtrack as before. Others, ‘ve develop children for that node ‘The running time ofthe algorithm depends on K and H,nbd Real Tins rocewor Py hat blevest be ost ikely te ale ose 9, that task by its destin “The RFB (Requests For Bids) c successfully execu ny proceso that can sucessfully exces he Frvceasor Py tating how quickly tan process he ontans the vital statistics ofthe tsk ang re the task t0 send @ bid 0 the fon sk. a a 1 sending processor estimates that there ‘An RFB is only sent ut i th Ha cncugh tine for timely reaponses toi Kalo computes the lates ine at whi foconed processor ean ofload the task onto a bidder without the tak deg being missed his time is given bythe expression “= "Task deadline (current time +time to move the ask + taskexecution ie) the an RFB is sent out 3 ft eecks to soe HH can meet the We ba Sta ‘When processor P, receives an RFB, requirements and sil execute its alteady-scheduled tasks successfully. Fr, si arrive and, how long if wil ake t0 be ei estimates when the new tak wil nrnted o fejected. This time i given by . 7 on = Curent time Hime forbid to be vessived by Py + time taken by P fo make a dession + time taken to transfer the task + time taken by Pt lther guarantee or rejest he tsk ‘Next it figures out the surplus computational time ‘yuries that it erent between gy, and the task deadtine D, This done by first estimating t compottiona time alteady spoken fo inthe interval [yp]. | time alloted to crite ws in [rx D] to run aready-aceapted noneiial sks + me needed in are + fiaction of ecetly accepted bids : time needed in f,,-D] 1 honor pending bids he G8 ie estimates. I'he task worstegee ei jose of Mecevative indeed IT aVERagS-case vals rg mee eh the bids wil ee ask execution time, hen no bid sent uty torts, outa, bid 10 Ps. The bid contains Phe 218K exection ee ne eh atin ong * - Fy mat td ns _ Ay bean review the BS es back tse wh eee ef sw ne ase "erative ean vary with he apt a 0 estimate of fon ih grated oe to proces the ts ch other proceso is most vn Ps computes the estimated arrival (ofthe asks the node Dent 5) and fy) Ly 86 yy VAS contained nh id ened og 1 () Ho computes the following quantity foreach i yy then. tai O~ Sraratas 1) spss) the maximum sch Yale. Then iP ot gn -|getak it ships the task out top, _|asae. The Buday strategy | the boddy strategy attacks the same typeof problem as the FAB algorithm. Sof el Te tasks arrive at the various process af a mle nd ia ital proceso finds ovo es offen sks es ay loaded processors. Te buy tae dis fom te FAB lerEred ond Realy U— (underloaded) ifQSTy itt, Mulile hardware units may be assigned odo the sme askin ogy ariel compu Ione oor of es unis n how pa + daogenes in he es. Te expt this to datetion. IF ony a minor ofthe wits ae fal and a mijorty of they proce the sme opt, We ean se this majority ret and ths me theefets offal. more than a minoriy ofthe wits disagree fre and ean use other methods, such 25 repeating the computation ser processors, to correct forthe faults. ‘& Conetion and masking are shor-erm measures, They newtalize fet ofthe observed flr 4 Inde second metho, is posible for systems 1 be desiged so tg spares can be switched into replace any ally un 1 Redundancy is expensive. Duplicating or tripling hardware i Tuxury thas only janie inthe most rita! aplicaions. Voting and consensus ‘One way of tsing redundancy isto have multiple unis exseue the sn i and compare thse outputs. I atleast thrce units are involved this compare ca ehoos the majority value a process alled voting and thus mask the effets of me flrs. vo unl re wed the comparison can detect (bu not cores) an exe 4} The designer must deside whether exact or approximae-apreemes i expected betwen funetonng units. - ‘© Approximate agreement may be used in enses where senso = smeasuting the pysieal enviroment © Thow ate tee types of voter, which can function in eases whet approximate agreements eqied 1. The formalized maory wee 2, The generalized K-pluaiy voter and 3. The gescraized median voter sileast we have deste | ca Btn . i software, SUPPOSE at itis reasonable gu 0 2 tio te ey cea om Spuiy.Ta eco MSY wo pn te ECORI AST aes, yp nen em og nye € Pyilf'd(s,y) $e, and Simei 2 Se it aaa frets tenga, Lee tr 2) semester enor xe eae nuratty veer "herd Kul oe vt dn eae a ester rent sy eres te a slong as P, contains atleast K elements oe ’ then tee exist some w« Puch ha \cineraized Median Voter ‘The generalized median voter works by selecting the middle vale. The idle nue i Selected by successively throwing away outlying values enti ely te nile vale is tf. sce Moor 48 a8 follows Let the cups eng vor on be es son ry) 1. Compute dy = d(x) forally €8 fori 4). 2 Let dy, be the maximum such d. Define S= S~ (x,,;)- 11'S contains only on element tht is the uit ofthe ote; else go back to stpEbel ond Real Fine stati Pairing ges and Stes iO whe "ely tansmitsiesnes mess nove th inert, ‘seen 10 be fay, jae, the IterFize Can checks tikew! 5 the output of the, homes ofall is to hardware processors in pais, ayy Es the some time. nd the itera i | the same way and produce exactly the same ony weg E Monitor and the inerfe wil il simultaneously °° 7 aneour EL poled 2 sc! rele robe that there sno failure ny oft fou conponen te processors the monitors and ihe ine, dell «gogo ts 1 vase we pa sin ant ale, Wen bln! nd ug nt one ata fe bs de aot Redundancy rep Nemo redunncy (NMR) a eh fr ral em teat yang proces ine one and vgn dupe 4 Nie suely od Folowing Fae ses is he N=. Om «atte approaches is psi In deg) te a ve ae tne case produce ouput Indesign) som ig, 420. Stale pring “The pir uns identical sofware using ena inputs a compres output of ech ask. “tate outputs a identical, ‘he pie detects nonidentleal outputs, “ofthe processors inthe pais fly. 4 The procesor that detects this discrepancy the est ofthe system, ths isolating his pair. approach wil work wells lng asthe fterface doesnot ila sors do nt fil identically and around the sine time first i not neces 1 point of fla the pair is functional. IF ether processor ‘hol isan indication that a est on x switches off he inet bth procs 4 The second isan acceptable assumption and the f ‘ccepable. To get around the inerfce beng 8 crite ‘we can introduce an interface monitor. Tree a Worn & we fs ve f—- ° » Fig, 422. Stectareof an NOIR laste (0) v0) ‘ig, 421 Use of monitor Single er