piecewise Liner SLAs and Performance Timetagment

Size: px
Start display at page:

Download "piecewise Liner SLAs and Performance Timetagment"

Transcription

1 i: Incrementl Cost bsed Scheduling under Piecewise Liner SLAs Yun Chi NEC Lbortories Americ 18 N. Wolfe Rd., SW3 35 Cupertino, CA 9514, USA lbs.com Hyun Jin Moon NEC Lbortories Americ 18 N. Wolfe Rd., SW3 35 Cupertino, CA 9514, USA lbs.com Hkn Hcıgümüş NEC Lbortories Americ 18 N. Wolfe Rd., SW3 35 Cupertino, CA 9514, USA lbs.com ABSTRACT In cloud computing environment, it is beneficil for the cloud service provider to offer differentited services mong different customers, who often hve different profiles. Therefore, -wre scheduling of queries is importnt. A prcticl -wre scheduling lgorithm must be ble to hndle the highly demnding query volumes in the scheduling queues to mke online scheduling decisions very quickly. We develop such highly efficient -wre query scheduling lgorithm, clled i. i tkes the query s derived from the service level greements (SLAs) between the service provider nd its customers into ccount to mke wre scheduling decisions. i is n incrementl vrition of n existing scheduling lgorithm,. Although exhibits n exceptionlly good performnce, it hs prohibitive complexity. Our min contributions re (1) to observe how behves under piecewise liner SLAs, which re very common in cloud computing systems, nd (2) to efficiently leverge these observtions nd to reduce the online complexity from O(N) for the originl version to O(log 2 N) for i. 1. INTRODUCTION In cloud computing environment, service providers offer vst IT resources to lrge sets of customers with diverse service requirements. By doing so, the service providers leverge the customer volume to chieve economies of scle. A service provider usully hs contrcts with its customers in the form of service level greements (SLAs), where the SLAs cn be bout mny spects of cloud computing service such s vilbility, security,, etc. An SLA indictes the level of service greed upon s well s the ssocited if the service provider fils to deliver the level of service. In this work, we focus on this defined by the SLAs. Obviously, optimizing the while serving lrge number of customers is vitl for the cloud service providers. Such n optimiztion hs to tke mny spects of cloud computing system into considertion, such s c- Permission to mke digitl or hrd copies of ll or prt of this work for personl or clssroom use is grnted without fee provided tht copies re not mde or distributed for profit or commercil dvntge nd tht copies ber this notice nd the full cittion on the first pge. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission nd/or fee. Articles from this volume were invited to present their results t The 37th Interntionl Conference on Very Lrge Dt Bses, August 29th September 3rd 211, Settle, Wshington. Proceedings of the VLDB Endowment, Vol. 4, No. 9 Copyright 211 VLDB Endowment /11/6... $ 1.. pcity plnning, disptching, nd scheduling. Therefore, single ll-inclusive solution is very difficult to obtin (see Appendix A for more discussions). Insted, in this pper we focus on locl optimiztion problem, nmely to schedule queries from diverse customers in -wre wy in order to minimize the SLA to the service provider. The problem of -wre scheduling hs been studied in the pst in the res of computer networks [12] nd rel dtbses [7]. However, it is somewht discourging to note tht the developed lgorithms hve not been widely deployed in rel pplictions, minly due to their high complexity. Such sitution seems to be chnging with the dvnce of the cloud computing. This is evident from the recent new interests in -wre scheduling lgorithms in the dtbse re [1, 4, 5]. We believe tht the cloud computing environment offers two resons for such chnge. Chnge in resource usge for scheduling: In order to mke scheduling decisions, we hve to use system resources (e.g., CPU cycles nd memory). In computer networks, for exmple, the rtio between the resources used by scheduling nd those used by processing network pcket hs to be very low, becuse of the limited resource cpcity t network router. This is min reson why simple policies such s firstcome-first-serve (FCFS), shortest-job-first (SJF), nd erliest-dedline-first (EDF) were preferred. In comprison, this rtio of resource usges is very different in the cloud computing environment. In the cloud computing environment, higher resource usge for scheduling is cceptble, s long s doing so is justified by the reduction in the query execution. Avilbility of functions: In trditionl computer networks nd rel- dtbses, there is lck of consensus on wht the most resonble function should be. In the cloud computing environment, such functions re usully vilble from the contrcts between the service providers nd their customers, in the form of service level greements (SLAs). In ddition, the SLAs used by cloud service providers re usully piecewise liner functions becuse they cn be defined by the cluses in business contrcts. Such piecewise liner SLAs, s we will show, mke mny -wre scheduling computtions more trctble. Bsed on the bove motivtions, we investigte wre scheduling lgorithm,, tht hs been proposed by Peh nd Tobgi [1, 12]. The min reson for us to choose 563

2 is its exceptionlly good performnce, which often pproches the lower bound of theoreticlly optiml totl [1]. Our own experiments, s will be shown lter, verified this clim, showing tht outperforms severl recently proposed -wre scheduling lgorithms in the dtbse systems [4] nd supercomputing [8] communities. However, suffers from one wek point its complexity for mking ech scheduling decision is O(N), where N is the number of queries to be scheduled. Such liner complexity is prohibitive in cloud computing system, where there cn be very lrge number of queries to be scheduled nd where the online decisions hve to be mde extremely quickly. We will discuss in detil in the next section. Here we only point out tht in, the priorities of the queries re dynmiclly chnging over. On one hnd, such dynmiclly chnging priorities cpture the chnge of urgency of queries over (e.g., s query pproching its dedline) thereby mking superior to lgorithms tht simply ssign sttic priority scores to the queries. On the other hnd, becuse the query priorities hve to be re-computed ech scheduling decision is mde, it seems tht the liner complexity O(N) is necessry price tht we hve to py. However, in this pper, we demonstrte tht the liner complexity of cn be voided, under piecewise liner functions (SLAs). The min contributions of this pper re two-fold. First, we observe tht in, lthough the priority scores of the queries chnge continuously over, the reltive order mong the priority scores only chnges s discrete events. In other words, the reltive order chnges over s series of discrete snpshots. Our second contribution is to develop n efficient scheduling lgorithm, i, to exploit the bove observtion. In i, the bove snpshots re incrementlly mintined, where from these snpshots, the query with the top priority cn be obtined very quickly. Our i lgorithm mkes exctly the sme scheduling decisions s. But compred with the O(N) complexity of, i chieves complexity of O(logN) for mny piecewise liner SLAs, nd O(log 2 N) for ll piecewise liner SLAs, by using techniques borrowed from the field of computtionl geometry. i is lso used s the scheduler in our comprehensive dt mngement pltform in the cloud, CloudDB [6]. 2. BACKGROUND AND RELATED WORK In this section, we provide bckground informtion. We first discuss SLAs, especilly vrious piecewise liner SLAs. Then we survey recent work on SLA-bsed scheduling lgorithms. Finlly, we introduce the lgorithm in detil. 2.1 Service Level Agreements (SLAs) A service level greement (SLA) is contrct between service provider nd its customers. An SLA indictes the level of service greed upon s well s the ssocited if the service provider fils to deliver the level of service. SLAs cnbeboutmnyspects ofcloudcomputingservicesuch s vilbility, security,, etc. In this pper, we focus on SLAs bout query. Tht is, n SLA is function f(t), where t is the of query nd f(t) is the corresponding if the is t. Fig. 1 shows severl exmples of SLAs used in previous work (e.g., [1, 4, 8]). Fig. 1 shows step-shped function, where the to the service provider is c 1 if the s1 (c) s1 t2 c2 Figure 1: Exmples of SLAs used in previous work. query is lter thn its dedline t 1. Fig. 1 shows function tht increses proportionlly to the trdiness, fter dedline t 1 is missed. In Fig. 1(c), the hs bound c 2, which indictes the mximum query cn introduce. It is worth noting tht lthough the SLAs shown in Fig. 1 represent mpping between system-level metrics(e.g., query ) nd s, the system-level performnce cn be drmticlly different from the performnce. For exmple, ssume service provider offers services to two customers, gold customer nd silver customer, by using shred dtbse server. Becuse the customers hve different SLAs, simply from the distribution of query for ll the queries, we re not ble to tell how good the performnce is in terms of. This is becuse the ltter depends on other fctors such s the shpe of the SLAs from the gold nd silver customers, the worklod rtios between the two customers, how the scheduling priority is determined, nd so on. Insted of functions, SLAs in some previous work re in the form of profit functions. We show in Appendix B.1 tht these two re equivlent. In ddition, in this pper we use per-query SLAs insted of per-customer or quntile SLAs, becuse we believe the former hve more flexibility. For exmple, under per-query SLAs, queries from the sme customer my hve different SLAs, depending on the expected execution nd resource consumption of the query, or depending on the budget quot tht the customer hs used so fr. In ddition, there exists techniques (e.g., [3]) tht directly mp per-query SLAs to quntile SLAs. 2.2 Piecewise Liner SLAs We focus on piecewise liner SLAs, s they exhibit very good model for cpturing service contrcts. By piecewise liner SLA, we men tht the function f(t) cn be divided into finite segments long the line nd in ech segment, f(t) is function liner in t. We next provide severl representtive piecewise liner SLAs nd show tht these SLAs re ble to cpture vrious rich semntics. Fig. 2 shows severl exmples of piecewise liner SLAs. In the figure, the x-xis represents query (for esy illustrtion, we ssume the query rrives to the system t ); the y-xis represents the corresponding to different t. We discuss these cses in detil. Fig. 2 describes n SLA with linerly incresing versus the. Tht is, the is proportionl to the. Such n SLA reflects trget of weighted men. A moment of thought will revel tht under such SLAs, minimizing the totl mong ll queries is equivlent to minimizing the weighted of ll queries, where the weight for ech query is proportionl to the slope of its linerly incresing SLA. Fig. 2 describes n SLA in the form of step function, where the is c if the is less thn t 1 nd c 1 otherwise. As specil cse, when c nd c 1 1 for ll queries, the verge under such n SLA is the sme 564

3 c c s (d) s1 c c s (e) c2 c c2 c s s1 (c) t2 (f) s2 t2 Figure 2: Representtive piecewise liner SLAs. s the frction of queries tht miss their (sole) dedlines. It is worth noting tht in this cse, nd in ll other cses in the figure, the prmeters such s c, c 1, nd t 1 cn be different for ech individul query. Tht is, ech query cn nme its own dedline nd corresponding for missing tht dedline. Fig. 2(c) describes stircse-shped SLA. The is c if the is less thn t 1, c 1 if the is between t 1 nd t 2, nd so on. Finlly, the becomes fixed vlue (c 2 in this exmple) fter the lst dedline (t 2 in this exmple). Such n SLA cn be used to cpture multiple semntics simultneously. For exmple, when checking out shopping crt t n online e-commerce Web site, less thn t 1 my result in good user experience wheres longer thn t 2 my reflect certin uncceptble performnce (e.g., t 2 cn be the out threshold of the user s browser, fter which the query result becomes invlid). Fig. 2(d) describes the mixture of step function nd liner function. Tht is, the remins constnt up to t 1 nd then grows linerly fterwrd. This SLA llows grce period, up to t 1, fter which the grows linerly over. This exmple lso illustrtes tht in generl, SLAs my contin jumps (t t 1 in this exmple) in the functions. Fig. 2(e) describes nother wy of mixing the step nd liner functions. The initilly grows linerly, but fter t 1, it becomes constnt. This SLA cptures, for exmple, the concept of proportionl with -out. Tht is, fter t 1, the dmge hs been done nd so the hs reched its mximl penlty vlue. Fig. 2(f) describes piecewise liner SLA in the most generl form. Tht is, the slope t ech segment cn be different nd there cn be jumps between consecutive segments in the SLA. Other thn cpturing vrious rich semntics, piecewise liner SLAs lso mke mny computtions in -wre scheduling more trctble, which is key to our work. 2.3 Relted Work Scheduling is mture problem nd hs long been extensively studied in vrious res such s computer networks, dtbse systems, nd Web services. There exist mny different scheduling policies nd there is vst mount of theoreticl results nd prcticl nlysis on vrious policies (e.g., see [9]). However, in most existing work on scheduling, the performnce metrics re low-level metrics (e.g., verge query or stretch [5]). Insted, in this subsection, we minly survey some of the recently proposed wre scheduling strtegies. As will be seen, this previous work ssumes some specific SLA shpes nd designs scheduling policies ccordingly, nd very often, the SLA shpes re piecewise liner. Guirguis et l. [4] im to minimize trdiness s shown in Fig. 2(d), which is i) zero if query is finished before its dedline, or ii) the query s finish minus its dedline, if the query misses it dedline. They nlyze tht EDF is better choice t minimizing trdiness when the system lod is low, nd SRPT(Shortest Remining Processing Time) is better choice when the lod is high. They propose ASETS*, novel scheduling method tht combines EDF nd SRPT using heuristic, nd show tht ASETS* outperforms both EDF nd SRPT over wide rnge of lods. Irwin et l. [8] consider SLA functions in the form of grdully incresing penlty with bound, s shown in Fig. 2(e) nd Fig. 1(c). In order to mximize service provider s profit, they employ two economic concepts, nmely present vlue nd opportunity, nd propose scheduling policy clled FirstRewrd tht combines the two concepts. Irwin et l. focus on the cse of preemptive execution, where query cn be stopped in the middle of execution nd resumed lter, which is often the cse in the supercomputing context. Note tht we ssume non-preemptive execution in this pper, which is often the cse in dt-intensive computing, such s dtbse query execution. Lstly, note tht FirstRewrd s scheduling decision tkes O(N 2 ), where N is the number of queries in the queue, nd this my cuse significnt performnce overhed. Peh [11] proposes scheduling nd dmission control method by using priority token bnk. It is ssumed tht jobs cn be grouped into N clsses nd jobs in ech clss re treted eqully (by using FIFO). Becuse different clsses hve different profiles (e.g., gurnteed or best-effort, different s for missing dedlines, etc.), prioritiztion scheme is designed to choose the next clss to serve. This method ssumes job clsses (videos, emils, etc.) nd does not differentite jobs within the sme clss. Therefore it is much more restricted thn the frmework, in which ech query cn hve its own profile. In[1], we study stircse-shped SLAs s shown in Fig. 2(c). We build dt structure, clled SLA-tree, to cpture the SLAs of the queries in the queue nd propose greedy lgorithm to improve given scheduling lgorithm. However, the SLA-tree bsed scheduling lgorithm in [1] cn only be used to improve n existing scheduling lgorithm nd it cn only hndle stircse-shped SLAs. While this previous work focuses on specific types of SLA function, (which we will introduce in detil momentrily) supports ll types of SLA functions. Also, our i supports ll piecewise liner SLA functions, which is much more expressive thn the functions in most of the previous work. We will compre the performnce of i with tht of some previous work mentioned bove in the experiment section. 2.4 Cost bsed Scheduling () in Detil is -bsed scheduling lgorithm proposed by Peh nd Tobgi [1, 12]. The gol of is to schedule queries in the queue, where ech query hs its own SLA in terms of, in wy tht minimizes the totl expected. However, chieving such miniml, even for the offline cse where future rrivls of queries re known beforehnd, is n NP-complete problem. The heuris- 565

4 tic used by is to compute priority score for ech query q in the queue bsed only on (1) q s SLA nd (2) how long q hs been witing in the queue. Becuse the priority score of q is computed independently from other queries in the queue, nd becuse the query with the highest vlue for the computed priority is chosen for execution, hs complexity of O(N) for mking scheduling decision when there re N queries in the queue. More specificlly, for ech query q in the queue, computes n expected reduction of executing q immeditely, rther thn delying it further. The expected reduction t given t (where t is the when the system schedules the next query to execute) is computed by r(t) e τ f(t t +τ)dτ f(t t ) (1) where t is the t which the query rrives to the system nd is free prmeter (the only prmeter) to be set by. For convenience, in the rest of the pper, we refer to the expected reduction r(t) computed by t t s the score t t. To intuitively describe how works, in Fig. 3 we illustrte how the score is computed t t nd t for query q with SLA of Fig. 2, nd we ssume the query rrives to the system t t. Fig. 3 shows how the score t t is obtined. At t, if q is served immeditely, the will be f(t t ) c. If, on the other hnd, we decide to postpone q, then it is ssumed tht q will sty in the queue for n mount of witing τ tht follows n exponentil distribution with men 1/. So the expected for postponing q is e τ f(t t +τ)dτ. The score is the net gin of choosing to serve q now insted of postponing it further 1. t e τ c t t t+ t t c t e τ t+ Figure 3: The intuition of : how scores re computed t t nd t t. So s cn be seen from Fig. 3, there re two fctors tht ffect the score of q. The first fctor is the SLA of q, i.e., the vlues of c, c 1, nd t 1; the second fctor is how long q hs been witing in the system. This second fctor cn be illustrted by compring Fig. 3 with Fig. 3. Compred with the sitution t t, t t, q hs been witing in the system longer, nd its dedline for the next jump in the SLA (the jump t t +t 1) is more urgent. As result, its score, which is reflected by the shded res in the figures, hs grown higher. 1 In the originl formul [12], the query execution q s ofqueryq is usedtoderivethe rel urgency ofq (i.e., the when the execution of q hs to be strted). In ddition, thescore ofq isr(t)/q s instedofr(t). Forconvenience, without loss of generlity, we describe s if q s were, nd use r(t) insted of r(t)/q s. In Appendix B.2, we show tht we cn compenste these differences by shifting nd scling the SLA of q. Some dditionl discussions bout re given in Appendix B, in which (1) we offer more detils on, (2) we prove tht for, without loss of generlity we cn lwys set c in piecewise liner SLAs, nd (3) we discuss two fctors, the multi-progrmming-level (MPL) nd the errors in the estimtion of query execution, tht my ffect the effectiveness of. From Eqution (1) we cn see tht is inefficient in tht t ech of scheduling, score hs to be computed for ech query in the queue nd ech computtion involves n integrtion. Assuming there re N queries witing in the queue nd ssuming no new query rrives, the totl number of scores to be computed in order to finish the N queries one by one is N 2 /2. As consequence, the verge for online scheduling of ech query is O(N). To llevite such expensive computtion of, in the next two sections, we propose n efficient incrementl version of, nmely i, tht hs O(log 2 N) complexity under piecewise liner SLAs. 3. I WITH O(logN) COMPLEXITY We develop our incrementl vrition of, i, in this section nd the next. For certin specil forms of piecewise liner SLAs, nmely those described in Figs. 2 2(d), i is ble to chieve n O(logN) complexity (compred with O(N) for originl ). We postpone i for the generl piecewise liner SLAs, nmely those described in Figs. 2(e) nd 2(f), to the next section. 3.1 Score without Integrtion We strt by showing method to compute the scores under piecewise liner SLAs without conducting ny integrtion. The method is bsed on cnonicl decomposition of piecewise liner SLAs nd on rules of clculus. We provide the derivtion detils in Appendix C nd give the finl formul, for the SLA described in Fig. 2(f), s the following r(t) 1 s s+(j1+s1 )e ( t) s2 s1 +(j 2+ )e (t 2 t) (2) where s, s 1, nd s 2 re the slopes of the SLA segments; j 1 nd j 2 re the heights of the jumps t t 1 nd t 2. Note tht Eqution (2) is vlid only for t < t 1. Therefore, when SLAs re piecewise liner, we cn efficiently compute scores t ny t by evluting the exponentil vlues t the plces where SLA segments chnge, without conducting ny integrtion. 3.2 i for the SLA in Figure 2 It cn be esily shown tht in, if the SLA is simple liner function, like tht in Fig. 2, then the score r(t) is constnt over t. The reson is tht by reshuffling Eqution (1), we hve r(t) e τ [f(t t +τ) f(t t )]dτ e τ bτdτ b/ where b is the slope of the liner SLA. So s cn be seen, in such cse, ech query hs -invrint score nd so incrementl cn be trivilly implemented by using priority queue, with n O(log N) complexity. 566

5 3.3 i for the SLA in Figure 2 To hndle the SLA in Fig. 2, we first derive tht for step-function SLA with height c 1, the score is score score r(t) c 1 e (t 1 t) e t where c 1 e t 1, nd t is the when the score is computed. Note tht the bove formul is vlid only for t < t 1 nd fter t 1, r(t) becomes. We cn see tht in this cse r(t) is -vrying. However, key observtion is tht does not cre the bsolute vlue for ech individul r(t), but insted cres bout the reltive order mong the r(t) s for ll the queries. This is becuse tht only hs to pick the top query with the highest score to be executed next. Therefore s long s the scores mong ll queries shre the sme fctor e t, their reltive order is -invrint ndonlydependsont, whenqueryrrives, t 1, when the dedline is, nd c 1, the of missing the query s dedline (recll tht without loss of generlity, we ssume c ). Therefore, this cse of step-function SLA cn be hndled by using priority queue ordered by the vlues of the queries (where vlues re -invrint), gin with n O(log N) complexity. 3.4 i for the SLA in Figure 2(c) Before hndlingslasin Figs. 2(c) nd2(d), we firststudy more generl cse s described in Fig. 4. Note tht in this more generl cse, the SLA is before t +t 1 nd n rbitrry function (which we denote by blob) fterwrd. Nowwestudyhowthe scorechngesover. Itturns out tht between t nd t +t 1, the score chnges in prticulr pttern. To see this, we compute the scores t two instnces, t nd t, where t < t < t < t +t 1. At t, we hve r(t ) e τ f(t t +τ)dτ f(t t ) nd t t we hve (derivtion given in Appendix B.5) r(t ) e (t t ) r(t ) (3) As consequence we cn see tht the score for the SLA in Fig. 4 grows exponentilly with prmeter until t +t 1, s shown in Fig. 4. t t t+ blob t score t t+ Figure 4: An SLA with zero until t +t 1 nd how the score chnges over exponentilly between t nd t +t 1. Now we come bck to the piecewise liner SLA described in Fig. 2(c). It cn be shown tht its score chnges over s described in Fig. 5. The bsic ide is tht t ech segment of the SLA, we cn shift down the stircses to get n SLA similr to tht in Fig. 4, where the blob is replced by the remining stircses. As elpses, s long s we re still in the sme segment, the score grows in n exponentil fshion, s we hve just shown. In ddition, t2 Figure 5: Left: how the score chnges over for the SLA in Fig. 2(c). Right: how the score chnges over for the SLA in Fig. 2(d). the exponentil rte is the sme for ll the segments nd for ll the queries, becuse the rte only depends on. Bsed on ll these properties, we develop n incrementl by using two priority queues. Both priority queues keep ll the queries, but with different priority scores. In the first priority queue, the score for ech query q is the score for q if we consider the remining stircses of q s blob. Following the sme ide s shown in the cse of SLA in Fig. 2, the reltive order mong the queries in the first priority queue remins fixed. So the problem is boiled down to detecting when ech query chnges its SLA segments (nd so chnges the shpe of the blob). For this, we use the second priority queue. In the second priority queue, the score for ech query q is the when the next jump will hppen in q s SLA, or in other words, the when q s current SLA segment will expire. At ech t when we need to select the next query to execute, we first tke out ll those queries whose SLA segments expired before t. For these queries, the old blobs re not vlid nymore nd so for ech such query q, we locte q s current SLA segment tht overlps t, get q s new of-expirtion of current SLA segment, nd reinsert q into the second priority queue. At the sme, we updte q s score in the first priority queue to reflect its new score. This second priority queue ctully mkes sure tht ll the queries re in their first segment of (probbly revised, if the old first segments hve expired) piecewise liner SLAs. Tht is, for ll queries, t < t 1 is lwys true in r(t), nd therefore Eqution (2) is lwys vlid. As cn be esily shown, the numbers of de-queue nd en-queue opertions for query q in the first nd second priority queues re both bounded by the number of segments in q s SLA, which we ssume to be bounded by constnt. So the mortized complexity is still O(log N) for scheduling ech query. 3.5 i for the SLA in Figure 2(d) For SLAs in Fig. 2(d), it cn be shown tht its score chnges over s described in Fig. 5, where initilly it grows exponentilly with rte nd fter t 1, it becomes constnt. So we cn use the similr ide s before, except tht we need three priority queues. The first nd the second ones re the sme s before, except when query is popped out from the second queue, it is never inserted bck; insted, it is removed from both the first nd the second priority queues nd is inserted into the third priority queue. This is becuse t the of being popped out of the second priority queue, q hs no further SLA segment chnges nd so its score becomes -invrint. When scheduling the next query to execute, fter updting queries in the first nd second priority queues, we compre the query with the highest score in the first priority queue nd tht in the third queue, nd pick the one with the higher score to execute. The complexity clerly is still O(log N). 567

6 4. I WITH O(log 2 N) COMPLEXITY In this section, we investigte the incrementl version of for generl piecewise liner SLAs, s described in Figs. 2(e) nd 2(f). We will nlyze why these cses re more difficult thn the previous cses. Then we describe n efficient method by using dynmic convex hull lgorithm in computtionl geometry. 4.1 Why the Difficult Cses Are Difficult Strting from Eqution (2), with certin derivtion we cn show tht the score t t for query q, who hs generl piecewise liner SLA, cn be written s f(t) α+e t, (4) where α nd re constnt vlues totlly determined by q s rrivl nd its SLA. It is obvious from Eqution (4) why n incrementl version of is hrd to get. Fig. 6 illustrtes for three smple queries with different α s nd s, how their scores chnge over. As cn be seen from Fig. 6, the reltive order of the scores of the three queries chnges over. In other words, in generl, the order mong the scores of the queries is different depending on the t when we sk bout tht order. Such property poses difficulties to n incrementl version of in the spce. To ddress such difficulties, in the following we investigte different view of the problem: we study the problem in the dul spce of liner functions. score q 3 α 2 q 2 q α Figure 6: How the SLA scores of three fictitious queries chnge over. The three queries in the dul spce, with the ffine-lines for scheduling. 4.2 A Dul Spce View of Now we provide new perspective of viewing the score f(t) α+e t. First, we define newvrible ξ e t (where ξ cn be considered s wrped line) to get α 3 α f(ξ) α+ξ, (5) where ξ is -vrying but α nd re constnts. Note tht there is one-to-one mpping between ξ nd t. Then, we put ech of these score functions into the dul spce of liner functions. Tht is, ech function f(ξ) α + ξ is mpped to point in the dul spce with coordinte (α,), s shown in Fig. 6. Next, we exmine t given t, how to determine the order mong the scores. With ξ e t, t t, we cn tret ξ s fixed nd look for the query tht mximizes f(ξ ) α + ξ. If we consider α nd s vribles, then in the dul spce, those points (α,) tht stisfy f(ξ ) α + ξ c re on n ffine line α c ξ, s shown in Fig. 6. So t given t, here is wy we cn get the point in the dul spce tht corresponds to the query with the highest score t t. We first get ξ e t -ξ nd construct n ffine line with slope ξ. Then strting from the upper-right corner of the dul spce, we move the ffine line towrd the lower-left corner of the dul spce while keeping the slope fixed t ξ. The first point hit by the moving ffine line will be the query tht hs the highest score t t. Figs. 7 nd 7 show the sme set of three queries, whose scores re compred t two different, t (with ξ e t ) nd t (with ξ e t ), wheret < t. Ascnbeseen, terlyt, the(bsolute) slope of the ffine line is smller nd so the point (α 3, 3) hs the highest score; s going, t t, (α 1, 1) hs the highest score becuse it hs higher vlue. α α 3 α 2 -ξ α α α 3 α 2 α Figure 7: Selecting the query with the top score t t nd t t. In summry, the dvntge of this dul-spce pproch is tht in the dul spce, we fix the (α,) loction of ech query s -invrint; nd over the we chnge the slope of the scheduling ffine line to select the query with the highest score t ny given moment. This dulspce perspective, s we show in Appendix D.1, lso offers insights bout why the cses in Section 3, i.e., those SLAs in Figs. 2 2(d), cn be hndled with O(logN). 4.3 A Convex Hull bsed Algorithm In this subsection, we develop i under generl piecewise liner SLAs tht tkes dvntges of this dul-spce view. The first question is whether we cn efficiently mintin the position of query q in the dul spce over. The nswer is ffirmtive, s the position of q in the dul spce cn only chnge finite number of s (which turns out to be bounded by the number of segments in q s SLA). We formlly prove this in Appendix D.2. Second, it cn be seen from Fig. 8 tht necessry condition for point q to be hit first by the moving ffine line is tht q is on the (upper) convex hull of ll points in the dul spce. Furthermore, it cn be shown tht if the convex hull is mintined properly, t ny t, finding the point hit first by the ffine line of slope -ξ (i.e., with ξ e t ) cn be done in O(logM), where M is the number of points on the convex hull (obviously, M N). More specificlly, if the points of the convex hull re sorted by their vlues, s shown in Fig. 8, then the ngles of the edges on the convex hull re in n monotoniclly decresing order. At t, we hve fixed slope -ξ, nd the condition for q on the convex hull to be the point hit first by the ffine line of slope -ξ is tht the slope -ξ is between the ngles of the two edges tht incident to q. This cn esily be done by conducting binry serch on the convex hull. Therefore, it follows tht we cn develop i s long s we cn (1) efficiently detect when query chnges its segment in its SLA nd (2) incrementlly mintin convex hull tht supports dding nd deleting points dynmiclly. For item (1), s discussed before, we cn mintin priority -ξ 568

7 α -ξ Figure 8: The queries with the highest score cn only be on the upper convex hull in the dul spce. The ngles of the edges on the convex hull follow monotonic decresing order. queue of queries ordered by the of the expirtion of the current SLA segment. At ny t, we cn quickly locte those queries whose current SLA segments hve expired before t, nd updte their new current SLA segments together with new expirtion of the current SLA segments. For item (2), we dopt, from the field of computtionl geometry, dynmic convex hull lgorithm for points in 2-D plne tht hs O(log 2 N) complexity [13]. The detiled implementtion of the i lgorithm, s well s the nd spce complexity nlysis, re given in Appendix D.3. An intuitive interprettion of i is tht it incrementlly mintins (1) the snpshot of the scores of ll the queries in the dul spce nd (2) the convex hull of the snpshot. At ech t of mking scheduling decision, i quickly brings up-to-dte both the snpshot nd the convex hull, from which the query with the top score t t cn be locted very efficiently. 5. EXPERIMENTAL STUDIES The experiments re composed of two sets. First, we verify the good performnce of (i 2 ), in terms of optimiztion, by compring it with severl other stte-ofthe-rt -wre scheduling lgorithms. Second, we compre the running of i with tht of the originl under vrious worklods. Some dditionl results re given in Appendix E. 5.1 Scheduling Effectiveness Verifiction In this subsection, we verify the performnce of (i) using two different types of SLA functions. We compre i with two pieces of previous work, ASETS* [4] nd FirstRewrd [8], s described t the beginning of this pper. For the query execution we use n exponentil distribution with men µ 3ms; for the rrivl rte we use Poisson rrivl with rte 1/λ, which is determined by µ nd the required lod. The dt set contins 2K queries, where the first 1K re used for system wrm-up. The first SLA function is shown in Fig. 1, which is lso known s weighted trdiness [4]. We set the dedline t 1 for ech query s 1 s its execution. Following the experiment design by [8], we choose the slope vlue from one of two clsses, high nd low, where hlf of queries get high slope vlues nd hlf get low slope vlues. Within ech clss, we choose the specific slope vlue from norml 2 In terms of the prmeter used in, Peh [1] showed tht the performnce is not sensitive to the exct vlue of s long s is in resonble rnge. So following Peh s suggestion, we simply used 1µ where µ is the men execution of query. α distribution with men nd stddev.2*men. The rtio between highmen ndlowmen is controlled bydecy skew fctor, which we vry in the experiment. Note tht ASETS* [4] is explicitly designed for this first SLA function. The second SLA function is shown in Fig. 1(c). Ech query incurs fter the minimum t 1, nd there is bound t 1 s the execution (t 2), which differentites it from the weighted trdiness bove. The bound vlue is chosen from two norml distributions s described bove, nd the rtio of high men nd low men is clled vlue skew fctor. Note tht FirstRewrd [8] is explicitly designed for this second SLA function. Tble 1(left) shows the result on the first SLA function under different decy skew fctors s described bove. When the system lod is.8, the three -wre scheduling lgorithms lredy significntly outperform the unwre counterprts, nmely FCFS nd SJF. The improvement become more striking when the system lod is incresed to.95. Among the three -wre scheduling lgorithms, i clerly outperforms the other two, especilly whenthesystemlod ndthedecyskewfctor rehigh(exctly when differentited services re needed the most). For exmple, under system lod of.95 nd decy skew fctor of 128, i outperforms the second best lgorithm by 45%. Tble 1(right) shows the result on the second SLA function under different vlued skew fctors s described bove. (We do not compre the performnce with ASETS*, becuse ASETS* cnnot hndle this SLA.) As cn be seen, lthough the mrgin is smller, i still consistently outperforms FirstRewrd, which ws designed specificlly for this second SLA, by bout 1%. 5.2 Scheduling Efficiency Study In this subsection, we investigte the performnce of i in terms of running 3. As for the bseline, for fir comprison, we implement so tht it computes scores under piecewise liner SLAs without conducting integrtion, s described in Section 3.1. All the running s re reported on Xeon PC with 3GHz CPU nd 4GB memory, running Fedor 11 Linux. When the system hs light lod, the number of queries to be scheduled is very smll in most of the with exponentilly distributed execution s. Therefore, in this experiment we focus on the cse when the system lod is.95, to study the sitution when the system is borderline overloded. Fig. 9 shows the running vs. the number of queries to be scheduled for nd i. As cn be seen, while exhibits cler O(N) behvior, the online scheduling of i is lmost unffected by N. It is worth noting tht FirstRewrd on verge needs couple of seconds to mke ech scheduling decision, which mkes it imprcticl for mny pplictions. Finlly, we study cse tht emultes the scenrio with mixture worklod with OLAP nd OLTP queries. Query execution in such cse cn be cptured by Preto (e.g., long-til) distribution. Following [5], we set the minimum vlue of the Preto to be 1ms nd the Preto index to be 1. We use lod of.8 becuse even with such light lod, the number of queries to be scheduled cn grow very 3 We implemented nive version of incrementl convex hull insted of using the sophisticted version in [13], nd in our experiments it is never bottleneck. Detils re skipped due to the spce limit. 569

8 Tble 1: Averge per query, for different lgorithms under different skew nd lod fctors. SLA Type SLA-1 SLA-2 Worklod Skew Fctor FCFS SJF ASETS* F-REWARD i per scheduling (ns) x 1 4 SLA 1 (Exp) i 1 2 queue size per scheduling (ns) 15 x SLA 2 (Exp) 14 i queue size Figure 9: Averge scheduling per query for the Exponentil cse with Lod.95. lrge (becuse of OLAP queries my occupy the dtbse server for rther long ). Note tht for this experiment, FirstRewrd ws not ble to finish within resonble (hours). Fig. 1 shows the performnce comprison nd we cn drw similr conclusions s before in terms of the running of i. One observtion is tht in this cse, while the to mke scheduling decision for i is extremely fst (less thn.1ms), tht for the originl cn become tens of milliseconds, n overhed uncceptble in mny dtbse pplictions. per scheduling (ns) 16 x SLA 1 (Preto) 16 1 x SLA 2 (Preto) i i queue sizex 1 4 per scheduling (ns) queue sizex 1 4 Figure 1: Averge scheduling per query for the Preto cse with Lod CONCLUSION We investigted -bsed scheduling lgorithm,, which hs superior performnce in terms of SLA nd therefore hs gret potentils in cloud computing. The min focus of this pper ws to implement in more efficient wy, where the is determined on query by using piecewise liner SLAs. To chieve this gol, we developed n incrementl vrition of, clled i, tht gretly reduces the computtion during online scheduling. i cn hndle ny type of piecewise liner SLAs, which re very common in cloud computing contrcts, with very competitive complexity. 7. ACKNOWLEDGMENTS We thnk Jeffrey Nughton for the insightful discussions nd his contributions. 8. REFERENCES [1] Y. Chi, H. J. Moon, H. Hcigumus, nd J. Ttemur. SLA-tree: A frmework for efficiently supporting SLA-bsed decisions in cloud computing. In EDBT, pges , 211. [2] A. Gnpthi, H. A. Kuno, U. Dyl, J. L. Wiener, A. Fox, M. I. Jordn, nd D. A. Ptterson. Predicting multiple metrics for queries: Better decisions enbled by mchine lerning. In ICDE, pges , 29. [3] D. Gmch, S. Krompss, A. Scholz, M. Wimmer, nd A. Kemper. Adptive qulity of service mngement for enterprise services. TWEB, 2(1):1 46, 28. [4] S. Guirguis, M. A. Shrf, P. K. Chrysnthis, A. Lbrinidis, nd K. Pruhs. Adptive scheduling of web trnsctions. In ICDE, pges , 29. [5] C. Gupt, A. Meht, S. Wng, nd U. Dyl. Fir, effective, efficient nd differentited scheduling in n enterprise dt wrehouse. In EDBT, pges , 29. [6] H. Hcıgümüş, J. Ttemur, W. Hsiung, H. J. Moon, O. Po, A. Swires, Y. Chi, nd H. Jfrpour. CloudDB: One size fits ll revived. In IEEE World Congress on Services (SERVICES), pges , 21. [7] J. R. Hrits, M. J. Crey, nd M. Livny. Vlue-bsed scheduling in rel- dtbse systems. The VLDB Journl, 2: , April [8] D. E. Irwin, L. E. Grit, nd J. S. Chse. Blncing risk nd rewrd in mrket-bsed tsk service. In HPDC, pges , 24. [9] J. Leung, L. Kelly, nd J. H. Anderson. Hndbook of Scheduling: Algorithms, Models, nd Performnce Anlysis. CRC Press, Inc., 24. [1] J. M. Peh. Scheduling nd dropping lgorithms to support integrted services in pcket-switched networks. PhD thesis, Stnford University, [11] J. M. Peh. Scheduling nd dmission control for integrted-services networks: the priority token bnk. Computer Networks, 31(23-24): , [12] J. M. Peh nd F. A. Tobgi. A -bsed scheduling lgorithm to support integrted services. In INFOCOM, pges , [13] F. P. Preprt nd M. I. Shmos. Computtionl geometry: n introduction. Springer-Verlg New York, Inc., New York, NY, USA,

9 APPENDIX A. CLOUD COMPUTING ARCHITECTURE Fig. 11 shows severl components tht we believe re crucil in cloud computing system. A cpcity plnning component plns (offline) nd mnges (online) the repliction nd system resources; disptching component disptches queries to pproprite servers in -wre wy; loclly t ech server, -wre scheduling lgorithm mkes besteffort scheduling to minimize the SLA mong queries disptched to the server. queries rrive q5 q4 queries rrive q3 query buffer scheduling q2 q1 cpcity plnning repliction migrtion re provisioning... disptching S dtbse sever query execution complete Figure 11: Components in cloud system. Idelly, we prefer comprehensive optimiztion by considering ll the bove components in globl wy. However, we hve concerns bout such globl optimiztion in terms of its dptbility to quick chnges nd sclbility to lrge numbers of clients in the cloud. Therefore, we choose to use modulr pproch where locl optimiztion is mde in ech component, nd in this pper we only focus on the -wre scheduling component. B. MORE DETAILS ABOUT In this ppendix, we discuss more detils bout, including the equivlence between SLAs nd profit SLAs, more detils of, setting c in piecewise liner SLAs for, some fctors tht my ffect the performnce of, nd the detiled derivtion of Eqution (3). B.1 Cost SLAs vs. Profit SLAs In our pper, to be consistent with the terminology used in the originl lgorithm, insted of using query profit, we use query. These two re equivlent due to the following reson. As shown in Fig. 12, we cn decompose profit SLA into the difference between constnt profit SLA nd SLA. Tht is, we cn ssume for ech query the service providerobtinsprofitofg 1 upfront, ndpysbck certin depending on the query, where the mximl pybck is g 1 +p 1. So the totl profit mong ll queries is the difference between the totl profit up front nd the totl. From the point of view of scheduling lgorithm, the totl upfront profit cn be considered s constnt, depending on the shpe of SLAs nd the worklod, which is beyond the control of the scheduling lgorithm; the totl is the prt tht scheduling lgorithm cn control, by chnging the priority mong queries, where minimizing the totl is the finl objective. g1 p1 profit t2 g1 profit g1+p1 t2 S1 S2 Sm Figure 12: Equivlence of profit nd SLAs. B.2 More Detils bout Score r(t) There re severl things we wnt topoint out bout Eqution (1). First, the score r(t) is -vrying. This mkes sense becuse the urgency of query depends on how long the query hs been witing in the queue so fr nd s result, it depends on the t when the score is computed. Second, the f(t t ) term in r(t) is the for executing query q right now (i.e., t t). Third, inside the integrtion, the dummy vrible τ indictes the further dely for query q nd f(t t +τ) is the corresponding. In the originl, the score for query q is r(t) divided by q s, the execution of q. However, equivlently we cn scle the SLA of q by 1/q s nd therefore use r(t) computed from the scled SLA s the score. In ddition, for simplicity of discussion, in this pper we illustrte the exmples s if the execution q s were. A non-zero execution q s cn be hndled by shifting q s SLA to the left by n mount of q s. These djustments re ctully implemented in ll our lgorithms mentioned in this pper. B.3 Why We Cn Set c under One observtion tht cn be obtined from Eqution (1) is tht without loss of generlity, to compute score, we cn ssume f(t ). This is becuse (s shown in Fig. 13) r g(t) e τ g(t t +τ)dτ g(t t ) e τ [f(t t +τ)+c ]dτ [f(t t )+c ] r f (t) where the lst step relies on the fct tht e τ is probbility density function nd therefore e τ c dτ c. This result reflects key feture of, where the dmge done so fr cn be ignored. Becuse of this property, in this pper, without loss of generlity, we ssume f(t ), nd therefore ll the c s in Figs. 2 (f) cn be set to. f(t) t t g(t)f(t)+c Figure 13: The illustrtion of why without loss of generlity, we cn ssume f(t ) in ll the SLAs. B.4 Fctors Tht Affect There re mny fctors tht potentilly ffect the effectiveness of. One of the fctors is the multi-progrmminglevel (MPL) of the dtbse. However, should not be very sensitive to MPL, becuse the score of query q only depends on q s SLA function nd q s urgency. The former is not relted to MPL nd the ltter is pproximted in by using n exponentil distribution, which is not sensitive to the exct MPL of the dtbse. Our experimentl studies, skipped due to the spce limits, hve verified this conjecture. Another fctor tht ffects the effectiveness of is how ccurte we cn estimte the query execution. We rgue tht fully -wre scheduling lgorithm hs to tke query execution into considertion (in this sense EDF is not fully -wre, becuse it only considers query dedlines but not query execution ). Otherwise, the urgency c t t 571

10 of query is not ccurtely cptured. To estimte query execution, recently work (e.g., [2]) hs strted using mchine lerning techniques nd obtined encourging results. In ddition, our own investigtion (prtly reported in [1]) demonstrte tht cn tolerte errors in the estimtion of query execution to certin levels. Specificlly, when the stndrd devitions of the errors re.2 nd 1., the performnce degrded by 1% nd 3%, respectively. B.5 Derivtion of Eqution (3) Here we provide the derivtion of Eqution(3) nd discuss some technicl subtleties. nd We hve r(t ) r(t ) e τ [f(t t +τ) f(t t )]dτ e τ [f(t t +τ) f(t t )]dτ e τ [f(t t +τ +t t ) f(t t )]dτ e (t t ) e η [f(t t +η) f(t t )]dη t t e (t t ) r(t ) where in the derivtion we defined new dummy vrible η τ + t t nd in the lst step, we used the fct tht f(t t +η) f(t t ) for η t t. This feture of is ctully due to the memory-less property of the exponentil distribution e τ, nmely t ny given, the conditionl distribution of the remining witing τ for query in the queue is the sme s the originl distribution of the totl witing. In ddition, techniclly, the blob in Fig. 4 is not relly rbitrry: the function hs to mke the integrtion in the score finite. However, such condition is stisfied by ny piecewise liner function with finite segments. C. FAST COMPUTATION OF SCORE UNDER PIECEWISE LINEAR SLAS In this ppendix, we derive n efficient pproch for computing scores of queries with piecewise liner SLAs. This pproch, by pplying cnonicl decomposition of piecewise liner SLAs nd by using rules of clculus, voids the expensive integrtion in the computtion of scores. Suchn efficient pproch, otherthn mkingit fst tocompute scores, revels importnt insights bout under piecewise liner SLAs. These insights turn out to be key to i, our incrementl implementtion of. C.1 Decomposition of Piecewise Liner SLAs We strt with cnonicl decomposition of generl piecewise liner SLA. We illustrte the decomposition by using the SLA described in Fig. 2(f), which is reproduced on the left in Fig. 14 s f 1. As shown in Fig. 14, the SLA cn be prtitioned into three segments: tht before t 1, tht between t 1 nd t 2, nd tht fter t 2. Strting from f 1, we decompose the SLA into the sum of f 2, which is continuous function, nd f 3 nd f 4, which re step functions tht cpture the jumps in the s of the SLA. Next, we tke the derivtive of f 2 to get f 2, which in turn cn be decomposed into the sum of f 5, which is constnt function corresponding to the initil slope of the SLA, nd f 6 nd f 7, which re step functions tht cpture the jumps in the slopes of the SLA. (Note tht f 2(t) is not defined t t 1 nd t 2, which is fine becuse f 2(t) is continuous nd finite.) It cn be shown tht in Eqution (4), the α vlue of score for piecewise liner SLA is determined by f 5 in Fig. 14, nd the vlue is determined by f 3, f 4, f 6 nd f 7 in Fig. 14. f1 c2 s s1 t2 s2 j2 j1 f2 s2 s s1 derivtive j1 j2 t2 f3 t2 f4 + + t2 s2 s s1 s s1 s s2 s1 f 2 t2 f5 f6 f7 + + t2 Figure 14: The decomposition of generl piecewise liner SLA f 1 into the sum of f 2, f 3, nd f 4, wheres f 2, the derivtive of f 2, is further decomposed into the sum of f 5, f 6, nd f 7. C.2 Fst Computtion of the Score It turns out tht with the help of the bove cnonicl decomposition, we re ble to compute the scores under piecewise liner SLAs without ny integrtion. We derive this result by using the decomposition shown in Fig. 14. The score r(t) for t < t 1 is computed s r(t) e τ f 1(t t +τ)dτ f 1(t t ) (6) where without loss of generlity we set f 1(t ) to be. For the step function f 3(τ), it cn be esily shown tht e τ f 3(t t +τ)dτ j 1 e (t 1 t), where t 1 is the when f 3 jumps from to vlue of j 1. We cn obtin similr result for f 4. Becuse integrtion is liner opertion, we hve r(t) e τ f 1(t t +τ)dτ f 1(t t ) e τ f 2(t t +τ)dτ f 1(t t ) +j 1 e (t 1 t) +j 2 e (t 2 t) e τ [f 2(t t +τ) f 2(t t )]dτ +j 1 e (t 1 t) +j 2 e (t 2 t) where we used the fct tht f 1(t t ) f 2(t t ) for t < t 1. The min technique we use next is the well-known rule of integrtion by prts in clculus: g (τ)f(τ)dτ f(τ)g(τ) g(τ)f (τ)dτ. (8) (7) 572

11 Compring Eqution (8) with the integrtion prt of the score in Eqution (7), nd by setting g(τ) e τ which implies tht g (τ) e τ, we hve g (τ)[f 2(t t +τ) f 2(t t )]dτ g(τ)[f 2(t t +τ) f 2(t t )] 1 Now from f 5, we hve 1 g(τ)f 2(t t +τ)dτ e τ f 2(t t +τ)dτ. e τ f 5(t t +τ)dτ 1 s, nd it cn be esily shown tht from f 6 nd f 7, 1 1 e τ s1 s f 6(t t +τ)dτ e ( t), e τ f 7(t t +τ)dτ which give the finl result r(t) 1 s2 s1 e (t2 t), s s+(j1+s1 )e ( t) s2 s1 +(j 2+ )e (t2 t). D. MORE DETAILS ABOUT I In this ppendix, we provide more detils bout i, including insights obtined from the dul-spce view, the correctness of n incrementl mintennce in the dul spce, the detiled i lgorithm, nd its complexity nlysis. D.1 Why the Esy Cses Are Esy With the dul-spce interprettion, now we cn revisit the esy cses in Section 3 nd see why they re esy to hndle. The queries with SLA of Fig. 2 re locted in the dul spce on the α-xis, s shown in Fig. 15. So no mtter when the scores re checked (i.e., no mtter with wht slope we hit the points using n ffine line), the best one is lwys the one with the highest α vlue. Similrly, the SLAs of Figs. 2,(c), nd (d) re shown in Fig. 15, where they re ll on the -xis in the dul spce. In this cse, no mtter using wht slope, the ffine line lwys hits the one with the highest vlue. All these clerly illustrte tht for cses of SLAs in Figs. 2 (d), the scores re order-preserving over (except when n SLA segment expires), nd so they re esy to hndle. α 3 α 2 α 1 α Figure 15: The queries in α-stge nd in - stge in the dul spce. For ese of discussion, in the reminder of pper (ppendices), we refer to the queries in Fig. 15 s queries in α-stge nd those in Fig. 15 s queries in -stge. Along the sme line, for those queries tht re neither in α-stge nor in -stge (i.e., those with α nd ), we refer to them s in α-stge. α D.2 Incrementl Mintennce Is Prcticl We prove tht it is prcticl to incrementlly mintin the convex hull in the dul spce for queries with piecewise liner SLAs. The keyis thefollowing propertyof scores under piecewise liner SLAs. Property 1. Under piecewise liner SLAs, in the dul spce, query q is either in α-stge, or -stge, or αstge. Moreover, the stge of q does not chnge within the sme segment of the piecewise liner SLA of q. Proof. Here we give sketch of the proof bsed on the decomposition of piecewise liner SLAs. Fig. 16 shows tht generl piecewise liner SLA for query q cn be decomposed into liner prt nd blob prt, where the liner prt determines the α vlue nd the blob prt determines vlue of q. Therefore, we either hve α (nd so q is in -stge), (nd so q is in α-stge), or α nd (nd so q is in α-stge). In ddition, from the figure we cn esily see tht such decomposition remins fixed within the sme segment of the SLA of q. t2 t2 SLA liner prt + blob prt Figure 16: Decomposition of generl piecewise liner SLA into liner prt, which determines α, nd blob prt, which determines. With such property we cn see tht query q cn only chnge its stge nd hence its position in the dul spce (q cn, e.g., chnge from α-stge bck to α-stge.) for finite number of s. And this number is bounded by the number of segments in q s piecewise liner SLA. D.3 i Algorithm in Detil nd Its Complexity Anlysis Fig. 17 nd Fig. 18 give the pseudo code for the i lgorithm. At ech t of scheduling the next query to execute, lgorithm schedulenext() is clled with t, the current stmp, nd newq, those queries tht hve rrived before t but fter the lst when schedulenext() ws clled. In the first step of schedulenext(), lgorithm updte() is clled in order to incrementlly updte the internl dt structures. The internl dt structures mintined by i (see Fig. 18) include (1) -Q: priority query contining queries either in -stge or in α-stge, where for ech query q in -Q, its priority is the q.ǫ, the when the current SLA segment of q will expire, (2) α-q, -Q, nd α-q: priority queues for those queries in α-stge, -stge, nd α-stge, respectively, nd (3) CH: the convex hull in the dul spce for ll the queries in α-stge. When updte() is clled, it first detects those queries whose current SLA segments hve expired before t (line 1). For these queries, their α nd vlues (s well s stges) my hve chnged since lst updte() ws clled. So these queries re removed from the corresponding internl dt structures nd these queries re ppended to the end of newq so tht their α nd vlues will be recomputed (lines 2 7). Then for ll the queries in newq (both those newly rrived nd those tht need updtes), their α nd + 573

12 Algorithm schedulenext(t,newq) input: t, newly rrived queries newq output: q, the next query to execute 1: cll updte(t,newq); 2: ξ e t ; 3: q α α-q.top(); 4: q -Q.top(); 5: q α CH.get(-ξ); 6: if q α hs the highest score 7: q α-q.pop(); 8: else if q hs the highest score 9: q -Q.pop(); 1: remove q from -Q; 11: else 12: q q α ; 13: remove q α from α-q, CH, nd -Q; 14: return q; Figure 17: Implementtion of schedulenext(). vlues re computed, their stges re determined, their new current SLA segments nd expirtion re updted, nd they re inserted into the corresponding dt structures. After the internl dt structures hve been updted, schedulenext() checks q α, q, nd q α, the three queries tht hve the highest scores in the α-q, -Q, nd α- Q. To get q α nd q, it is suffice to peek t the tops of α-q nd -Q; to get q α, it is suffice to do binry serch on the convex hull CH by using the slope ξ e t. Then the best one mong q α, q, nd q α is selected nd removed from the corresponding dt structures(ssuming this query is ctully executed next). The nd spce complexities of the overll i re given s the following. Property 2. Our i implementtion hs n O(log 2 N) complexity for scheduling ech query, nd it hs n O(N) spce complexity. Algorithm updte(t,newq) input: t, newly rrived queries newq -Q: priority queue on expirtion (ǫ) α-q: priority queue for queries in α-stge -Q: priority queue for queries in -stge α-q: priority queue for queries in α-stge CH: convex hull for queries in α-stge First, hndle queries tht chnge SLA sections 1: forech q in -Q such tht q.ǫ < t do 2: remove q from -Q; 3: ppend q to the end of newq; 4: if q ws in -Q 5: remove q from -Q; 6: else if q ws in α-q 7: remove q from α-q nd CH; Second, (re)insert updted queries 8: forech q in newq do 9: compute q s α,, ǫ vlues; 1: if q is in α-stge 11: insert q to α-q; 12: else if q is in -stge 13: insert q to -Q nd -Q; 14: else /* q is in α-stge */ 15: insert q to α-q, CH, nd -Q; Figure 18: Implementtion of the updte(). Proof. In schedulenext(), lines 3,4,7,9, nd 1 ll hve complexity O(log N), by using n ddressble priority queue or blnced binry serch tree. Line 5 hs complexity O(logN) by using binry serch. Line 13 my involve the updte of the convex hull nd so hs complexity of O(log 2 N). For updte(), ech query q cn enter the loops of lines 1 7 nd lines 8 15 t most K s where K is constnt tht represents the number of segments in q s SLA. Ech step in the two loops hs n O(logN) complexity other thn lines 7 nd 15, in which it tkes O(log 2 N) to dd or remove q from the convex hull dynmiclly. In the bove nlysis, we ssumed K to be constnt. If this is not the cse, then the complexity becomes O(K log 2 N) for i nd O(KN) for. In terms of spce complexity, we notice tht ech query q occurs t most once in -Q nd once in either α-q, -Q, nd α-q, nd in ddition, the size of the convex hull CH is bounded by the size of α-q. So we cn see tht the spce complexity for i is O(N). E. ADDITIONAL EXPERIMENTS In this ppendix, we show the running of vs. i under generl piecewise liner SLAs. The SLAs re in the shpe tht is described in Fig. 2(f), nmely with three segments where s, s 1 nds 2 re non-zero. Fig. 19shows the performnce comprison between nd i, in terms of running under different queue sizes, for query execution following exponentil distribution (Fig. 19-left) nd Preto distribution (Fig. 19-right). Severl observtions cn be obtined from the results. First, the running for obviously scles linerly with respect to the queue size nd for i, the running is reltively insensitive to the queue size. Second, becuse there re more queries in α-stge, the running for i is higher thn tht in Fig. 9 nd hs higher vritions. This is more distinct in Fig. 19-left, where the running for is low due to smller queue sizes. However, even for Preto distribution, the number of queries in their α-stge is not very high (in tens). This is becuse when lrge OLAP query blocks the server, most queries will be in their lst SLA segment, i.e., in α-stge. In other words, the sophisticted incrementl convex hull pproch is not wrrnted in this cse nd the min benefit comes from queries in α nd stges. However, even this is the cse, the dulspce view is still crucil for obtining the benefit of i. per scheduling (ns) 2.5 x SLA 3 Exp 14 i queue size per scheduling (ns) 9 x SLA 3 Preto 15 8 i queue size Figure 19: Averge scheduling per query for the Exponentil (left) nd Preto (right) cses under generl piecewise liner SLAs, with Lod

All pay auctions with certain and uncertain prizes a comment

All pay auctions with certain and uncertain prizes a comment CENTER FOR RESEARC IN ECONOMICS AND MANAGEMENT CREAM Publiction No. 1-2015 All py uctions with certin nd uncertin prizes comment Christin Riis All py uctions with certin nd uncertin prizes comment Christin

More information

Reasoning to Solve Equations and Inequalities

Reasoning to Solve Equations and Inequalities Lesson4 Resoning to Solve Equtions nd Inequlities In erlier work in this unit, you modeled situtions with severl vriles nd equtions. For exmple, suppose you were given usiness plns for concert showing

More information

Graphs on Logarithmic and Semilogarithmic Paper

Graphs on Logarithmic and Semilogarithmic Paper 0CH_PHClter_TMSETE_ 3//00 :3 PM Pge Grphs on Logrithmic nd Semilogrithmic Pper OBJECTIVES When ou hve completed this chpter, ou should be ble to: Mke grphs on logrithmic nd semilogrithmic pper. Grph empiricl

More information

Econ 4721 Money and Banking Problem Set 2 Answer Key

Econ 4721 Money and Banking Problem Set 2 Answer Key Econ 472 Money nd Bnking Problem Set 2 Answer Key Problem (35 points) Consider n overlpping genertions model in which consumers live for two periods. The number of people born in ech genertion grows in

More information

ClearPeaks Customer Care Guide. Business as Usual (BaU) Services Peace of mind for your BI Investment

ClearPeaks Customer Care Guide. Business as Usual (BaU) Services Peace of mind for your BI Investment ClerPeks Customer Cre Guide Business s Usul (BU) Services Pece of mind for your BI Investment ClerPeks Customer Cre Business s Usul Services Tble of Contents 1. Overview...3 Benefits of Choosing ClerPeks

More information

Polynomial Functions. Polynomial functions in one variable can be written in expanded form as ( )

Polynomial Functions. Polynomial functions in one variable can be written in expanded form as ( ) Polynomil Functions Polynomil functions in one vrible cn be written in expnded form s n n 1 n 2 2 f x = x + x + x + + x + x+ n n 1 n 2 2 1 0 Exmples of polynomils in expnded form re nd 3 8 7 4 = 5 4 +

More information

LINEAR TRANSFORMATIONS AND THEIR REPRESENTING MATRICES

LINEAR TRANSFORMATIONS AND THEIR REPRESENTING MATRICES LINEAR TRANSFORMATIONS AND THEIR REPRESENTING MATRICES DAVID WEBB CONTENTS Liner trnsformtions 2 The representing mtrix of liner trnsformtion 3 3 An ppliction: reflections in the plne 6 4 The lgebr of

More information

9 CONTINUOUS DISTRIBUTIONS

9 CONTINUOUS DISTRIBUTIONS 9 CONTINUOUS DISTIBUTIONS A rndom vrible whose vlue my fll nywhere in rnge of vlues is continuous rndom vrible nd will be ssocited with some continuous distribution. Continuous distributions re to discrete

More information

Integration. 148 Chapter 7 Integration

Integration. 148 Chapter 7 Integration 48 Chpter 7 Integrtion 7 Integrtion t ech, by supposing tht during ech tenth of second the object is going t constnt speed Since the object initilly hs speed, we gin suppose it mintins this speed, but

More information

Factoring Polynomials

Factoring Polynomials Fctoring Polynomils Some definitions (not necessrily ll for secondry school mthemtics): A polynomil is the sum of one or more terms, in which ech term consists of product of constnt nd one or more vribles

More information

Use Geometry Expressions to create a more complex locus of points. Find evidence for equivalence using Geometry Expressions.

Use Geometry Expressions to create a more complex locus of points. Find evidence for equivalence using Geometry Expressions. Lerning Objectives Loci nd Conics Lesson 3: The Ellipse Level: Preclculus Time required: 120 minutes In this lesson, students will generlize their knowledge of the circle to the ellipse. The prmetric nd

More information

Helicopter Theme and Variations

Helicopter Theme and Variations Helicopter Theme nd Vritions Or, Some Experimentl Designs Employing Pper Helicopters Some possible explntory vribles re: Who drops the helicopter The length of the rotor bldes The height from which the

More information

How To Set Up A Network For Your Business

How To Set Up A Network For Your Business Why Network is n Essentil Productivity Tool for Any Smll Business TechAdvisory.org SME Reports sponsored by Effective technology is essentil for smll businesses looking to increse their productivity. Computer

More information

Basic Analysis of Autarky and Free Trade Models

Basic Analysis of Autarky and Free Trade Models Bsic Anlysis of Autrky nd Free Trde Models AUTARKY Autrky condition in prticulr commodity mrket refers to sitution in which country does not engge in ny trde in tht commodity with other countries. Consequently

More information

EQUATIONS OF LINES AND PLANES

EQUATIONS OF LINES AND PLANES EQUATIONS OF LINES AND PLANES MATH 195, SECTION 59 (VIPUL NAIK) Corresponding mteril in the ook: Section 12.5. Wht students should definitely get: Prmetric eqution of line given in point-direction nd twopoint

More information

Operations with Polynomials

Operations with Polynomials 38 Chpter P Prerequisites P.4 Opertions with Polynomils Wht you should lern: Write polynomils in stndrd form nd identify the leding coefficients nd degrees of polynomils Add nd subtrct polynomils Multiply

More information

How To Network A Smll Business

How To Network A Smll Business Why network is n essentil productivity tool for ny smll business Effective technology is essentil for smll businesses looking to increse the productivity of their people nd processes. Introducing technology

More information

PROF. BOYAN KOSTADINOV NEW YORK CITY COLLEGE OF TECHNOLOGY, CUNY

PROF. BOYAN KOSTADINOV NEW YORK CITY COLLEGE OF TECHNOLOGY, CUNY MAT 0630 INTERNET RESOURCES, REVIEW OF CONCEPTS AND COMMON MISTAKES PROF. BOYAN KOSTADINOV NEW YORK CITY COLLEGE OF TECHNOLOGY, CUNY Contents 1. ACT Compss Prctice Tests 1 2. Common Mistkes 2 3. Distributive

More information

Small Business Networking

Small Business Networking Why network is n essentil productivity tool for ny smll business Effective technology is essentil for smll businesses looking to increse the productivity of their people nd business. Introducing technology

More information

Small Business Networking

Small Business Networking Why network is n essentil productivity tool for ny smll business Effective technology is essentil for smll businesses looking to increse the productivity of their people nd processes. Introducing technology

More information

COMPARISON OF SOME METHODS TO FIT A MULTIPLICATIVE TARIFF STRUCTURE TO OBSERVED RISK DATA BY B. AJNE. Skandza, Stockholm ABSTRACT

COMPARISON OF SOME METHODS TO FIT A MULTIPLICATIVE TARIFF STRUCTURE TO OBSERVED RISK DATA BY B. AJNE. Skandza, Stockholm ABSTRACT COMPARISON OF SOME METHODS TO FIT A MULTIPLICATIVE TARIFF STRUCTURE TO OBSERVED RISK DATA BY B. AJNE Skndz, Stockholm ABSTRACT Three methods for fitting multiplictive models to observed, cross-clssified

More information

CHAPTER 11 Numerical Differentiation and Integration

CHAPTER 11 Numerical Differentiation and Integration CHAPTER 11 Numericl Differentition nd Integrtion Differentition nd integrtion re bsic mthemticl opertions with wide rnge of pplictions in mny res of science. It is therefore importnt to hve good methods

More information

Small Business Cloud Services

Small Business Cloud Services Smll Business Cloud Services Summry. We re thick in the midst of historic se-chnge in computing. Like the emergence of personl computers, grphicl user interfces, nd mobile devices, the cloud is lredy profoundly

More information

Treatment Spring Late Summer Fall 0.10 5.56 3.85 0.61 6.97 3.01 1.91 3.01 2.13 2.99 5.33 2.50 1.06 3.53 6.10 Mean = 1.33 Mean = 4.88 Mean = 3.

Treatment Spring Late Summer Fall 0.10 5.56 3.85 0.61 6.97 3.01 1.91 3.01 2.13 2.99 5.33 2.50 1.06 3.53 6.10 Mean = 1.33 Mean = 4.88 Mean = 3. The nlysis of vrince (ANOVA) Although the t-test is one of the most commonly used sttisticl hypothesis tests, it hs limittions. The mjor limittion is tht the t-test cn be used to compre the mens of only

More information

5.2. LINE INTEGRALS 265. Let us quickly review the kind of integrals we have studied so far before we introduce a new one.

5.2. LINE INTEGRALS 265. Let us quickly review the kind of integrals we have studied so far before we introduce a new one. 5.2. LINE INTEGRALS 265 5.2 Line Integrls 5.2.1 Introduction Let us quickly review the kind of integrls we hve studied so fr before we introduce new one. 1. Definite integrl. Given continuous rel-vlued

More information

Enterprise Risk Management Software Buyer s Guide

Enterprise Risk Management Software Buyer s Guide Enterprise Risk Mngement Softwre Buyer s Guide 1. Wht is Enterprise Risk Mngement? 2. Gols of n ERM Progrm 3. Why Implement ERM 4. Steps to Implementing Successful ERM Progrm 5. Key Performnce Indictors

More information

2. Transaction Cost Economics

2. Transaction Cost Economics 3 2. Trnsction Cost Economics Trnsctions Trnsctions Cn Cn Be Be Internl Internl or or Externl Externl n n Orgniztion Orgniztion Trnsctions Trnsctions occur occur whenever whenever good good or or service

More information

Small Business Networking

Small Business Networking Why network is n essentil productivity tool for ny smll business Effective technology is essentil for smll businesses looking to increse the productivity of their people nd business. Introducing technology

More information

Lecture 3 Gaussian Probability Distribution

Lecture 3 Gaussian Probability Distribution Lecture 3 Gussin Probbility Distribution Introduction l Gussin probbility distribution is perhps the most used distribution in ll of science. u lso clled bell shped curve or norml distribution l Unlike

More information

Small Business Networking

Small Business Networking Why network is n essentil productivity tool for ny smll business Effective technology is essentil for smll businesses looking to increse the productivity of their people nd processes. Introducing technology

More information

Babylonian Method of Computing the Square Root: Justifications Based on Fuzzy Techniques and on Computational Complexity

Babylonian Method of Computing the Square Root: Justifications Based on Fuzzy Techniques and on Computational Complexity Bbylonin Method of Computing the Squre Root: Justifictions Bsed on Fuzzy Techniques nd on Computtionl Complexity Olg Koshelev Deprtment of Mthemtics Eduction University of Texs t El Pso 500 W. University

More information

Protocol Analysis. 17-654/17-764 Analysis of Software Artifacts Kevin Bierhoff

Protocol Analysis. 17-654/17-764 Analysis of Software Artifacts Kevin Bierhoff Protocol Anlysis 17-654/17-764 Anlysis of Softwre Artifcts Kevin Bierhoff Tke-Awys Protocols define temporl ordering of events Cn often be cptured with stte mchines Protocol nlysis needs to py ttention

More information

Experiment 6: Friction

Experiment 6: Friction Experiment 6: Friction In previous lbs we studied Newton s lws in n idel setting, tht is, one where friction nd ir resistnce were ignored. However, from our everydy experience with motion, we know tht

More information

Data replication in mobile computing

Data replication in mobile computing Technicl Report, My 2010 Dt repliction in mobile computing Bchelor s Thesis in Electricl Engineering Rodrigo Christovm Pmplon HALMSTAD UNIVERSITY, IDE SCHOOL OF INFORMATION SCIENCE, COMPUTER AND ELECTRICAL

More information

Math 135 Circles and Completing the Square Examples

Math 135 Circles and Completing the Square Examples Mth 135 Circles nd Completing the Squre Exmples A perfect squre is number such tht = b 2 for some rel number b. Some exmples of perfect squres re 4 = 2 2, 16 = 4 2, 169 = 13 2. We wish to hve method for

More information

Algebra Review. How well do you remember your algebra?

Algebra Review. How well do you remember your algebra? Algebr Review How well do you remember your lgebr? 1 The Order of Opertions Wht do we men when we write + 4? If we multiply we get 6 nd dding 4 gives 10. But, if we dd + 4 = 7 first, then multiply by then

More information

Economics Letters 65 (1999) 9 15. macroeconomists. a b, Ruth A. Judson, Ann L. Owen. Received 11 December 1998; accepted 12 May 1999

Economics Letters 65 (1999) 9 15. macroeconomists. a b, Ruth A. Judson, Ann L. Owen. Received 11 December 1998; accepted 12 May 1999 Economics Letters 65 (1999) 9 15 Estimting dynmic pnel dt models: guide for q mcroeconomists b, * Ruth A. Judson, Ann L. Owen Federl Reserve Bord of Governors, 0th & C Sts., N.W. Wshington, D.C. 0551,

More information

g(y(a), y(b)) = o, B a y(a)+b b y(b)=c, Boundary Value Problems Lecture Notes to Accompany

g(y(a), y(b)) = o, B a y(a)+b b y(b)=c, Boundary Value Problems Lecture Notes to Accompany Lecture Notes to Accompny Scientific Computing An Introductory Survey Second Edition by Michel T Heth Boundry Vlue Problems Side conditions prescribing solution or derivtive vlues t specified points required

More information

Value Function Approximation using Multiple Aggregation for Multiattribute Resource Management

Value Function Approximation using Multiple Aggregation for Multiattribute Resource Management Journl of Mchine Lerning Reserch 9 (2008) 2079-2 Submitted 8/08; Published 0/08 Vlue Function Approximtion using Multiple Aggregtion for Multittribute Resource Mngement Abrhm George Wrren B. Powell Deprtment

More information

Section 7-4 Translation of Axes

Section 7-4 Translation of Axes 62 7 ADDITIONAL TOPICS IN ANALYTIC GEOMETRY Section 7-4 Trnsltion of Aes Trnsltion of Aes Stndrd Equtions of Trnslted Conics Grphing Equtions of the Form A 2 C 2 D E F 0 Finding Equtions of Conics In the

More information

Mathematics. Vectors. hsn.uk.net. Higher. Contents. Vectors 128 HSN23100

Mathematics. Vectors. hsn.uk.net. Higher. Contents. Vectors 128 HSN23100 hsn.uk.net Higher Mthemtics UNIT 3 OUTCOME 1 Vectors Contents Vectors 18 1 Vectors nd Sclrs 18 Components 18 3 Mgnitude 130 4 Equl Vectors 131 5 Addition nd Subtrction of Vectors 13 6 Multipliction by

More information

Decision Rule Extraction from Trained Neural Networks Using Rough Sets

Decision Rule Extraction from Trained Neural Networks Using Rough Sets Decision Rule Extrction from Trined Neurl Networks Using Rough Sets Alin Lzr nd Ishwr K. Sethi Vision nd Neurl Networks Lbortory Deprtment of Computer Science Wyne Stte University Detroit, MI 48 ABSTRACT

More information

Introducing Kashef for Application Monitoring

Introducing Kashef for Application Monitoring WextWise 2010 Introducing Kshef for Appliction The Cse for Rel-time monitoring of dtcenter helth is criticl IT process serving vriety of needs. Avilbility requirements of 6 nd 7 nines of tody SOA oriented

More information

Euler Euler Everywhere Using the Euler-Lagrange Equation to Solve Calculus of Variation Problems

Euler Euler Everywhere Using the Euler-Lagrange Equation to Solve Calculus of Variation Problems Euler Euler Everywhere Using the Euler-Lgrnge Eqution to Solve Clculus of Vrition Problems Jenine Smllwood Principles of Anlysis Professor Flschk My 12, 1998 1 1. Introduction Clculus of vritions is brnch

More information

Section 5-4 Trigonometric Functions

Section 5-4 Trigonometric Functions 5- Trigonometric Functions Section 5- Trigonometric Functions Definition of the Trigonometric Functions Clcultor Evlution of Trigonometric Functions Definition of the Trigonometric Functions Alternte Form

More information

GFI MilArchiver 6 vs C2C Archive One Policy Mnger GFI Softwre www.gfi.com GFI MilArchiver 6 vs C2C Archive One Policy Mnger GFI MilArchiver 6 C2C Archive One Policy Mnger Who we re Generl fetures Supports

More information

Techniques for Requirements Gathering and Definition. Kristian Persson Principal Product Specialist

Techniques for Requirements Gathering and Definition. Kristian Persson Principal Product Specialist Techniques for Requirements Gthering nd Definition Kristin Persson Principl Product Specilist Requirements Lifecycle Mngement Elicit nd define business/user requirements Vlidte requirements Anlyze requirements

More information

Example A rectangular box without lid is to be made from a square cardboard of sides 18 cm by cutting equal squares from each corner and then folding

Example A rectangular box without lid is to be made from a square cardboard of sides 18 cm by cutting equal squares from each corner and then folding 1 Exmple A rectngulr box without lid is to be mde from squre crdbord of sides 18 cm by cutting equl squres from ech corner nd then folding up the sides. 1 Exmple A rectngulr box without lid is to be mde

More information

ORBITAL MANEUVERS USING LOW-THRUST

ORBITAL MANEUVERS USING LOW-THRUST Proceedings of the 8th WSEAS Interntionl Conference on SIGNAL PROCESSING, ROBOICS nd AUOMAION ORBIAL MANEUVERS USING LOW-HRUS VIVIAN MARINS GOMES, ANONIO F. B. A. PRADO, HÉLIO KOII KUGA Ntionl Institute

More information

Vendor Rating for Service Desk Selection

Vendor Rating for Service Desk Selection Vendor Presented By DATE Using the scores of 0, 1, 2, or 3, plese rte the vendor's presenttion on how well they demonstrted the functionl requirements in the res below. Also consider how efficient nd functionl

More information

Appendix D: Completing the Square and the Quadratic Formula. In Appendix A, two special cases of expanding brackets were considered:

Appendix D: Completing the Square and the Quadratic Formula. In Appendix A, two special cases of expanding brackets were considered: Appendi D: Completing the Squre nd the Qudrtic Formul Fctoring qudrtic epressions such s: + 6 + 8 ws one of the topics introduced in Appendi C. Fctoring qudrtic epressions is useful skill tht cn help you

More information

The Velocity Factor of an Insulated Two-Wire Transmission Line

The Velocity Factor of an Insulated Two-Wire Transmission Line The Velocity Fctor of n Insulted Two-Wire Trnsmission Line Problem Kirk T. McDonld Joseph Henry Lbortories, Princeton University, Princeton, NJ 08544 Mrch 7, 008 Estimte the velocity fctor F = v/c nd the

More information

Integration by Substitution

Integration by Substitution Integrtion by Substitution Dr. Philippe B. Lvl Kennesw Stte University August, 8 Abstrct This hndout contins mteril on very importnt integrtion method clled integrtion by substitution. Substitution is

More information

2005-06 Second Term MAT2060B 1. Supplementary Notes 3 Interchange of Differentiation and Integration

2005-06 Second Term MAT2060B 1. Supplementary Notes 3 Interchange of Differentiation and Integration Source: http://www.mth.cuhk.edu.hk/~mt26/mt26b/notes/notes3.pdf 25-6 Second Term MAT26B 1 Supplementry Notes 3 Interchnge of Differentition nd Integrtion The theme of this course is bout vrious limiting

More information

Regular Sets and Expressions

Regular Sets and Expressions Regulr Sets nd Expressions Finite utomt re importnt in science, mthemtics, nd engineering. Engineers like them ecuse they re super models for circuits (And, since the dvent of VLSI systems sometimes finite

More information

Health insurance exchanges What to expect in 2014

Health insurance exchanges What to expect in 2014 Helth insurnce exchnges Wht to expect in 2014 33096CAEENABC 02/13 The bsics of exchnges As prt of the Affordble Cre Act (ACA or helth cre reform lw), strting in 2014 ALL Americns must hve minimum mount

More information

Vectors 2. 1. Recap of vectors

Vectors 2. 1. Recap of vectors Vectors 2. Recp of vectors Vectors re directed line segments - they cn be represented in component form or by direction nd mgnitude. We cn use trigonometry nd Pythgors theorem to switch between the forms

More information

QoS Mechanisms C HAPTER 3. 3.1 Introduction. 3.2 Classification

QoS Mechanisms C HAPTER 3. 3.1 Introduction. 3.2 Classification C HAPTER 3 QoS Mechnisms 3.1 Introduction In the previous chpter, we introduced the fundmentl QoS concepts. In this chpter we introduce number of key QoS mechnisms tht enble QoS services. At the end of

More information

Modeling POMDPs for Generating and Simulating Stock Investment Policies

Modeling POMDPs for Generating and Simulating Stock Investment Policies Modeling POMDPs for Generting nd Simulting Stock Investment Policies Augusto Cesr Espíndol Bff UNIRIO - Dep. Informátic Aplicd Av. Psteur, 458 - Térreo Rio de Jneiro - Brzil ugusto.bff@uniriotec.br Angelo

More information

9.3. The Scalar Product. Introduction. Prerequisites. Learning Outcomes

9.3. The Scalar Product. Introduction. Prerequisites. Learning Outcomes The Sclr Product 9.3 Introduction There re two kinds of multipliction involving vectors. The first is known s the sclr product or dot product. This is so-clled becuse when the sclr product of two vectors

More information

WEB DELAY ANALYSIS AND REDUCTION BY USING LOAD BALANCING OF A DNS-BASED WEB SERVER CLUSTER

WEB DELAY ANALYSIS AND REDUCTION BY USING LOAD BALANCING OF A DNS-BASED WEB SERVER CLUSTER Interntionl Journl of Computers nd Applictions, Vol. 9, No., 007 WEB DELAY ANALYSIS AND REDUCTION BY USING LOAD BALANCING OF A DNS-BASED WEB SERVER CLUSTER Y.W. Bi nd Y.C. Wu Abstrct Bsed on our survey

More information

Warm-up for Differential Calculus

Warm-up for Differential Calculus Summer Assignment Wrm-up for Differentil Clculus Who should complete this pcket? Students who hve completed Functions or Honors Functions nd will be tking Differentil Clculus in the fll of 015. Due Dte:

More information

ASG Techniques of Adaptivity

ASG Techniques of Adaptivity ASG Techniques of Adptivity Hrld Meyer nd Dominik Kuropk nd Peter Tröger Hsso-Plttner-Institute for IT-Systems-Engineering t the University of Potsdm Prof.-Dr.-Helmert-Strsse 2-3, 14482 Potsdm, Germny

More information

Lecture 5. Inner Product

Lecture 5. Inner Product Lecture 5 Inner Product Let us strt with the following problem. Given point P R nd line L R, how cn we find the point on the line closest to P? Answer: Drw line segment from P meeting the line in right

More information

Unleashing the Power of Cloud

Unleashing the Power of Cloud Unleshing the Power of Cloud A Joint White Pper by FusionLyer nd NetIQ Copyright 2015 FusionLyer, Inc. All rights reserved. No prt of this publiction my be reproduced, stored in retrievl system, or trnsmitted,

More information

TITLE THE PRINCIPLES OF COIN-TAP METHOD OF NON-DESTRUCTIVE TESTING

TITLE THE PRINCIPLES OF COIN-TAP METHOD OF NON-DESTRUCTIVE TESTING TITLE THE PRINCIPLES OF COIN-TAP METHOD OF NON-DESTRUCTIVE TESTING Sung Joon Kim*, Dong-Chul Che Kore Aerospce Reserch Institute, 45 Eoeun-Dong, Youseong-Gu, Dejeon, 35-333, Kore Phone : 82-42-86-231 FAX

More information

AntiSpyware Enterprise Module 8.5

AntiSpyware Enterprise Module 8.5 AntiSpywre Enterprise Module 8.5 Product Guide Aout the AntiSpywre Enterprise Module The McAfee AntiSpywre Enterprise Module 8.5 is n dd-on to the VirusScn Enterprise 8.5i product tht extends its ility

More information

P.3 Polynomials and Factoring. P.3 an 1. Polynomial STUDY TIP. Example 1 Writing Polynomials in Standard Form. What you should learn

P.3 Polynomials and Factoring. P.3 an 1. Polynomial STUDY TIP. Example 1 Writing Polynomials in Standard Form. What you should learn 33337_0P03.qp 2/27/06 24 9:3 AM Chpter P Pge 24 Prerequisites P.3 Polynomils nd Fctoring Wht you should lern Polynomils An lgeric epression is collection of vriles nd rel numers. The most common type of

More information

A Decision Theoretic Framework for Ranking using Implicit Feedback

A Decision Theoretic Framework for Ranking using Implicit Feedback A Decision Theoretic Frmework for Rnking using Implicit Feedbck Onno Zoeter Michel Tylor Ed Snelson John Guiver Nick Crswell Mrtin Szummer Microsoft Reserch Cmbridge 7 J J Thomson Avenue Cmbridge, United

More information

6.2 Volumes of Revolution: The Disk Method

6.2 Volumes of Revolution: The Disk Method mth ppliction: volumes of revolution, prt ii Volumes of Revolution: The Disk Method One of the simplest pplictions of integrtion (Theorem ) nd the ccumultion process is to determine so-clled volumes of

More information

Health insurance marketplace What to expect in 2014

Health insurance marketplace What to expect in 2014 Helth insurnce mrketplce Wht to expect in 2014 33096VAEENBVA 06/13 The bsics of the mrketplce As prt of the Affordble Cre Act (ACA or helth cre reform lw), strting in 2014 ALL Americns must hve minimum

More information

Week 7 - Perfect Competition and Monopoly

Week 7 - Perfect Competition and Monopoly Week 7 - Perfect Competition nd Monopoly Our im here is to compre the industry-wide response to chnges in demnd nd costs by monopolized industry nd by perfectly competitive one. We distinguish between

More information

JaERM Software-as-a-Solution Package

JaERM Software-as-a-Solution Package JERM Softwre-s--Solution Pckge Enterprise Risk Mngement ( ERM ) Public listed compnies nd orgnistions providing finncil services re required by Monetry Authority of Singpore ( MAS ) nd/or Singpore Stock

More information

DlNBVRGH + Sickness Absence Monitoring Report. Executive of the Council. Purpose of report

DlNBVRGH + Sickness Absence Monitoring Report. Executive of the Council. Purpose of report DlNBVRGH + + THE CITY OF EDINBURGH COUNCIL Sickness Absence Monitoring Report Executive of the Council 8fh My 4 I.I...3 Purpose of report This report quntifies the mount of working time lost s result of

More information

Redistributing the Gains from Trade through Non-linear. Lump-sum Transfers

Redistributing the Gains from Trade through Non-linear. Lump-sum Transfers Redistributing the Gins from Trde through Non-liner Lump-sum Trnsfers Ysukzu Ichino Fculty of Economics, Konn University April 21, 214 Abstrct I exmine lump-sum trnsfer rules to redistribute the gins from

More information

VoIP for the Small Business

VoIP for the Small Business Reducing your telecommunictions costs VoIP (Voice over Internet Protocol) offers low cost lterntive to expensive trditionl phone services nd is rpidly becoming the communictions system of choice for smll

More information

Implementation Evaluation Modeling of Selecting ERP Software Based on Fuzzy Theory

Implementation Evaluation Modeling of Selecting ERP Software Based on Fuzzy Theory Implementtion Evlution Modeling of Selecting ERP Softwre Bsed on Fuzzy Theory Xukn Xu 1, Ydong Jing 1 nd Zheng Shi 2 1 Informtion Mngement Deprtment, Chngzhou Cmpus, Hohi University, Chngzhou 213022, Jingsu,

More information

4.11 Inner Product Spaces

4.11 Inner Product Spaces 314 CHAPTER 4 Vector Spces 9. A mtrix of the form 0 0 b c 0 d 0 0 e 0 f g 0 h 0 cnnot be invertible. 10. A mtrix of the form bc d e f ghi such tht e bd = 0 cnnot be invertible. 4.11 Inner Product Spces

More information

Distributions. (corresponding to the cumulative distribution function for the discrete case).

Distributions. (corresponding to the cumulative distribution function for the discrete case). Distributions Recll tht n integrble function f : R [,] such tht R f()d = is clled probbility density function (pdf). The distribution function for the pdf is given by F() = (corresponding to the cumultive

More information

Project 6 Aircraft static stability and control

Project 6 Aircraft static stability and control Project 6 Aircrft sttic stbility nd control The min objective of the project No. 6 is to compute the chrcteristics of the ircrft sttic stbility nd control chrcteristics in the pitch nd roll chnnel. The

More information

An Undergraduate Curriculum Evaluation with the Analytic Hierarchy Process

An Undergraduate Curriculum Evaluation with the Analytic Hierarchy Process An Undergrdute Curriculum Evlution with the Anlytic Hierrchy Process Les Frir Jessic O. Mtson Jck E. Mtson Deprtment of Industril Engineering P.O. Box 870288 University of Albm Tuscloos, AL. 35487 Abstrct

More information

Efficient load-balancing routing for wireless mesh networks

Efficient load-balancing routing for wireless mesh networks Computer Networks 51 (007) 50 66 www.elsevier.com/locte/comnet Efficient lod-blncing routing for wireless mesh networks Yigl Bejerno, Seung-Je Hn b, *,1, Amit Kumr c Bell Lbortories, Lucent Technologies,

More information

Object Semantics. 6.170 Lecture 2

Object Semantics. 6.170 Lecture 2 Object Semntics 6.170 Lecture 2 The objectives of this lecture re to: to help you become fmilir with the bsic runtime mechnism common to ll object-oriented lnguges (but with prticulr focus on Jv): vribles,

More information

Hillsborough Township Public Schools Mathematics Department Computer Programming 1

Hillsborough Township Public Schools Mathematics Department Computer Programming 1 Essentil Unit 1 Introduction to Progrmming Pcing: 15 dys Common Unit Test Wht re the ethicl implictions for ming in tody s world? There re ethicl responsibilities to consider when writing computer s. Citizenship,

More information

Application Bundles & Data Plans

Application Bundles & Data Plans Appliction Appliction Bundles & Dt Plns We ve got plns for you. Trnsporttion compnies tody ren t one-size-fits-ll. Your fleet s budget, size nd opertions re unique. To meet the needs of your fleet nd help

More information

On the Robustness of Most Probable Explanations

On the Robustness of Most Probable Explanations On the Robustness of Most Probble Explntions Hei Chn School of Electricl Engineering nd Computer Science Oregon Stte University Corvllis, OR 97330 chnhe@eecs.oregonstte.edu Adnn Drwiche Computer Science

More information

Learning to Search Better than Your Teacher

Learning to Search Better than Your Teacher Ki-Wei Chng University of Illinois t Urbn Chmpign, IL Akshy Krishnmurthy Crnegie Mellon University, Pittsburgh, PA Alekh Agrwl Microsoft Reserch, New York, NY Hl Dumé III University of Mrylnd, College

More information

Automated Resource Management in Virtualized Data Centers

Automated Resource Management in Virtualized Data Centers Automted Resource Mngement in Virtulized Dt Centers Shrd Singhl Xioyun Zhu Hewlett Pcrd Lbortories VMwre Inc. IM 9 Tutoril: Recent Advnces in the Appliction of Control Theory to Networ nd Service Mngement

More information

Example 27.1 Draw a Venn diagram to show the relationship between counting numbers, whole numbers, integers, and rational numbers.

Example 27.1 Draw a Venn diagram to show the relationship between counting numbers, whole numbers, integers, and rational numbers. 2 Rtionl Numbers Integers such s 5 were importnt when solving the eqution x+5 = 0. In similr wy, frctions re importnt for solving equtions like 2x = 1. Wht bout equtions like 2x + 1 = 0? Equtions of this

More information

Performance analysis model for big data applications in cloud computing

Performance analysis model for big data applications in cloud computing Butist Villlpndo et l. Journl of Cloud Computing: Advnces, Systems nd Applictions 2014, 3:19 RESEARCH Performnce nlysis model for big dt pplictions in cloud computing Luis Edurdo Butist Villlpndo 1,2,

More information

Network Configuration Independence Mechanism

Network Configuration Independence Mechanism 3GPP TSG SA WG3 Security S3#19 S3-010323 3-6 July, 2001 Newbury, UK Source: Title: Document for: AT&T Wireless Network Configurtion Independence Mechnism Approvl 1 Introduction During the lst S3 meeting

More information

and thus, they are similar. If k = 3 then the Jordan form of both matrices is

and thus, they are similar. If k = 3 then the Jordan form of both matrices is Homework ssignment 11 Section 7. pp. 249-25 Exercise 1. Let N 1 nd N 2 be nilpotent mtrices over the field F. Prove tht N 1 nd N 2 re similr if nd only if they hve the sme miniml polynomil. Solution: If

More information

Solving BAMO Problems

Solving BAMO Problems Solving BAMO Problems Tom Dvis tomrdvis@erthlink.net http://www.geometer.org/mthcircles Februry 20, 2000 Abstrct Strtegies for solving problems in the BAMO contest (the By Are Mthemticl Olympid). Only

More information

DEVELOPMENT. Introduction to Virtualization E-book. anow is the time to realize all of the benefits of virtualizing your test and development lab.

DEVELOPMENT. Introduction to Virtualization E-book. anow is the time to realize all of the benefits of virtualizing your test and development lab. Introduction to Virtuliztion E-book S Now is the time to relize ll of the benefits of virtulizing your test nd development lb. YOUR CHAPTER 3 p 2 A TEST AND p 4 VOLATILE IT S p 7 p 9 p 10 YOUR CHAPTER

More information

Universal Regularizers For Robust Sparse Coding and Modeling

Universal Regularizers For Robust Sparse Coding and Modeling 1 Universl Regulrizers For Robust Sprse Coding nd Modeling Igncio Rmírez nd Guillermo Spiro Deprtment of Electricl nd Computer Engineering University of Minnesot {rmir48,guille}@umn.edu Abstrct rxiv:13.2941v2

More information

Bayesian Updating with Continuous Priors Class 13, 18.05, Spring 2014 Jeremy Orloff and Jonathan Bloom

Bayesian Updating with Continuous Priors Class 13, 18.05, Spring 2014 Jeremy Orloff and Jonathan Bloom Byesin Updting with Continuous Priors Clss 3, 8.05, Spring 04 Jeremy Orloff nd Jonthn Bloom Lerning Gols. Understnd prmeterized fmily of distriutions s representing continuous rnge of hypotheses for the

More information

The Black-Litterman Model For Active Portfolio Management Forthcoming in Journal of Portfolio Management Winter 2009

The Black-Litterman Model For Active Portfolio Management Forthcoming in Journal of Portfolio Management Winter 2009 The lck-littermn Model For Active Portfolio Mngement Forthcoming in Journl of Portfolio Mngement Winter 009 Alexndre Schutel D Silv Senior Vice President, Quntittive Investment Group Neuberger ermn Wi

More information

How To Study The Effects Of Music Composition On Children

How To Study The Effects Of Music Composition On Children C-crcs Cognitive - Counselling Reserch & Conference Services (eissn: 2301-2358) Volume I Effects of Music Composition Intervention on Elementry School Children b M. Hogenes, B. Vn Oers, R. F. W. Diekstr,

More information

Online Multicommodity Routing with Time Windows

Online Multicommodity Routing with Time Windows Konrd-Zuse-Zentrum für Informtionstechnik Berlin Tkustrße 7 D-14195 Berlin-Dhlem Germny TOBIAS HARKS 1 STEFAN HEINZ MARC E. PFETSCH TJARK VREDEVELD 2 Online Multicommodity Routing with Time Windows 1 Institute

More information

Portfolio approach to information technology security resource allocation decisions

Portfolio approach to information technology security resource allocation decisions Portfolio pproch to informtion technology security resource lloction decisions Shivrj Knungo Deprtment of Decision Sciences The George Wshington University Wshington DC 20052 knungo@gwu.edu Abstrct This

More information