A Distributed Dynamic Load Balancer for Iterative Applications
|
|
|
- Rafe Richard
- 10 years ago
- Views:
Transcription
1 A Distributed Dyamic Balacer for Iterative Applicatios Harshitha Meo, Laxmikat Kalé Departmet of Computer Sciece, Uiversity of Illiois at Urbaa-Champaig ABSTRACT For may applicatios, computatio load varies over time. Such applicatios require dyamic load balacig to improve performace. Cetralized load balacig schemes, which perform the load balacig decisios at a cetral locatio, are ot scalable. I cotrast, fully distributed strategies are scalable but typically do ot produce a balaced work distributio as they ted to cosider oly local iformatio. This paper describes a fully distributed algorithm for load balacig that uses partial iformatio about the global state of the system to perform load balacig. This algorithm, referred to as GrapevieLB, cosists of two stages: global iformatio propagatio usig a lightweight algorithm ispired by epidemic [2] algorithms, ad work uit trasfer usig a radomized algorithm. We provide aalysis of the algorithm alog with detailed simulatio ad performace compariso with other load balacig strategies. We demostrate the effectiveess of GrapevieLB for adaptive mesh refiemet ad molecular dyamics o up to 3,72 cores of BlueGee/Q. Geeral Terms Algorithms, Performace Keywords load balacig, distributed load balacer, epidemic algorithm. INTRODUCTION imbalace is a isidious factor that ca reduce the performace of a parallel applicatio sigificatly. For some applicatios, such as basic stecil codes for structured grids, the load is easy to predict ad does ot vary dyamically. However, for a sigificat class of applicatios, load represeted by pieces of computatios varies over time, ad may be harder to predict. This is becomig icreasigly prevalet with the emergece of sophisticated applicatios. Permissio to make digital or hard copies of all or part of this work for persoal or classroom use is grated without fee provided that copies are ot made or distributed for profit or commercial advatage ad that copies bear this otice ad the full citatio o the first page. Copyrights for compoets of this work owed by others tha ACM must be hoored. Abstractig with credit is permitted. To copy otherwise, or republish, to post o servers or to redistribute to lists, requires prior specific permissio ad/or a fee. Request permissios from [email protected]. SC 3, November 7-2, 23, Dever, Colorado, USA Copyright 23 ACM /3/...$5.. For example, atoms movig i a molecular dyamics simulatio will lead to (almost) o imbalace whe they are distributed statically to processors. But, they create imbalace whe spatial partitioig of atoms is performed for more sophisticated ad efficiet force evaluatio algorithms. The presece of moisture ad clouds i weather simulatios, elemets turig from elastic to plastic i structural dyamics simulatios ad dyamic adaptive mesh refiemets are all examples of sophisticated applicatios which have a strog tedecy for load imbalace. All the examples above are of iterative applicatios: the program executes series of time-steps, or iteratios, leadig to covergece of some error metric. Cosecutive iteratios have relatively similar patters of commuicatio ad computatio. There is aother class of applicatios, such as combiatorial search, that ivolves dyamic creatio of work ad therefore has a tedecy for imbalace. This class of applicatios has distict characteristics ad load balacig eeds, ad has bee addressed by much past work such as work-stealig [25, 3, 32]. This paper does ot focus o such applicatios but istead o the iterative applicatios, which are predomiat i sciece ad egieerig. We also do ot focus o approaches that partitio the fie-graied applicatio data. For example, i ustructured mesh based applicatios, the etire mesh (cosistig of billios of elemets) may be partitioed by a library such as METIS [3]. This approach is expesive ad ot widely applicable; istead we focus o scearios where the applicatio work has already bee partitioed ito coarser work uits. For iterative applicatios, the basic scheme we follow is: The applicatio is assumed to cosist of a large umber of migratable uits (for example, these could be chuks of meshes i adaptive mesh refiemet applicatio). The applicatio pauses after every so may iteratios, ad the load balacer decides whether to migrate some of these uits to restore balace. balacig is expesive i these scearios ad is performed ifrequetly or wheever sigificat imbalace is detected. Note that a reactive strategy such as work-stealig, which is triggered whe a processor is idle, is almost ifeasible (e.g. Commuicatio to existig tasks must be redirected o the fly). Schemes for arrivig at a distributed cosesus o whe ad how ofte to balace load [28], ad how to avoid the pause (carryig out load balacig asychroously with the applicatio) have bee addressed i the past. I this paper we focus o a sychroous load balacer. Sice scietific applicatios have sychroizatios at various poits, this ca be used without extra overhead of sychroizatio.
2 Various strategies have bee proposed to address the load balacig problem. May applicatios employ cetralized load balacig strategies, where load iformatio is collected o to a sigle processor, ad their decisio algorithm is ru sequetially. Such strategies have bee show to be effective for a few hudred to thousad processors, because the total umber of work uits is relatively small (o the order of te to hudred per processor). However, they preset a clear performace bottleeck beyod a few thousad processors, ad may become ifeasible due to the memory capacity bottleeck o a sigle processor. A alterative to cetralized strategies are distributed strategies that use local iformatio, e.g. diffusio based [9]. I a distributed strategy, each processor makes autoomous decisios based o its local view of the system. The local view typically cosists of the load of its eighborig processors. Such strategies are scalable, but ted to yield poor load balace due to the limited local iformatio []. Hierarchical strategies [35, 23, ] overcome some of the aforemetioed disadvatages. They create subgroups of processors, ad collect iformatio at the root of each subgroup. Higher levels i the hierarchy oly receive aggregate iformatio ad deliver decisios i aggregate terms. Although effective i reducig memory costs, ad esurig good balace, these strategies may suffer from excessive data collectio at the lowest level of the hierarchy ad work beig doe at multiple levels. We propose a fully distributed strategy, GrapevieLB, that has bee desiged to overcome the drawback of other distributed strategies by obtaiig a partial represetatio of the global state of the system ad basig the load balacig decisios o this. We describe a light weight iformatio propagatio algorithm based o epidemic algorithm [2](also kow as the gossip protocol []) to propagate the load iformatio about the uderloaded processors i the system to the overloaded processors. This spreads the iformatio i the same fashio as gossip spreads through the grapevie i a society. Based o this iformatio, GrapevieLB makes probabilistic trasfer of work uits to obtai good load distributio. The proposed algorithm is scalable ad ca be tued to optimize for either cost or performace. The primary cotributios of this paper are: GrapevieLB, a fully distributed load balacig algorithm that attais a load balacig quality comparable to the cetralized strategies while icurrig sigificatly less overhead. Aalysis of propagatio algorithm used by GrapevieLB which leads us to a iterestig observatio that good load balace ca be achieved with sigificatly less iformatio about uderloaded processors i the system. Detailed evaluatios that experimetally demostrate the scalability ad quality of GrapevieLB usig simulatio. Demostratio of its effectiveess i compariso to several other load balacig strategies for adaptive mesh refiemet ad molecular dyamics o up to 3,72 cores of BlueGee/Q. 2. BACKGROUND characteristics i dyamic applicatios ca chage Processor L avg L max σ I p p2 p3 p4 p5 p6 p p2 p3 p4 p5 p Table : Choice of load imbalace metric over time. Therefore, such applicatios require periodic load balacig to maitai good system utilizatio. To eable load balacig, a popular approach is overdecompositio. The applicatio writer exposes parallelism by overdecomposig the computatio ito tasks or objects. The problem is decomposed ito commuicatig objects ad the ru-time system ca assig these objects to processors ad perform rebalacig. The load balacig problem i our cotext ca be summarized as: give a distributed collectio of work uits, each with a load estimate, decide which work uits should be moved to which processors, to reduce the load imbalace. The load balacer eeds iformatio about the loads preseted by each work-uit. This ca be based o a model (simple examples beig associatig a fixed amout of work with each grid poit, or particle). But for may applicatios, aother metric turs out to be more accurate. For these applicatios, a heuristic called priciple of persistece [2] holds which allows us to use recet istrumeted history as a guide to predictig load i ear-future iteratios. The load balacig strategy we describe ca be used with either model-based or persistece-based load predictios. I persistece-based load balacer, the statistics about the load of each task o a processor is collected at that processor. The database cotaiig the task iformatio is used by the load balacers to produce a ew mappig. The ru-time system the migrates the tasks based o this mappig. It is importat to choose the right metric to quatify load imbalace i the system. Usig stadard deviatio to measure load imbalace may seem like a appropriate metric, but cosider the two scearios show i Table. I both the cases, the average load of the system is 2. If we cosider stadard deviatio, σ, to be a measure of imbalace, the we fid that i case ad case 2 we obtai the same σ of 6 whereas the utilizatio ad the total applicatio times differ. A better idicator of load imbalace i the system is the ratio of maximum load to average load. More formally, load imbalace (I) ca be measured usig I = Lmax L avg () I case, I is.5 ad i case 2, I is. We use this metric of load imbalace as oe of the evaluatio criteria to measure the performace of the load balacig strategy. Notice that this criteria is domiated by the load of the sigle processor viz. the most overloaded processor because of the max operator. This is correct, sice the executio time is determied by the worst-loaded processor ad others must wait for it to complete its step. Apart from how well the load balacer ca balace load,
3 it is importat to icur less overhead due to load balacig. Otherwise, the beefit of load balacig is lost i the overhead. Therefore, we evaluate quality of load balace, cost of the load balacig strategy ad the total applicatio time. 3. RELATED WORK balacig has bee studied extesively i the literature. For applicatios with regular load, static load balacig ca be performed where load balace is achieved by carefully mappig the data oto processors. Numerous algorithms have bee developed for statically partitioig a computatioal mesh [2, 5, 6, 6]. These model the computatio as a graph ad use graph partitioig algorithms to divide the graph amog processors. Graph ad hypergraph partitioig techiques have bee used to map tasks o to processors to balace load while cosiderig the locality. They are geerally used as a pre-processig step ad ted to be expesive. Our algorithm is employed where the applicatio work has already bee partitioed ad used to balace the computatio load imbalace that arises as the applicatio progresses. Our algorithm also takes ito cosideratio the existig mappig ad moves tasks oly if a processor is overloaded. For irregular applicatios, work stealig is employed i task schedulig ad is part of rutime systems such as Cilk [3]. Work stealig is traditioally used for task parallelism of the kid see i combiatorial search or dividead-coquer applicatios, where tasks are beig geerated cotiuously. A recet work by Dia et al. [] scales work stealig to 892 processors usig the PGAS programmig model ad RDMA. I work that followed, a hierarchical techique described as retetive work stealig was employed to scale work-stealig to over 5K cores by exploitig the priciple of persistece to iteratively refie the load balace of task-based applicatios [23]. CHAOS [3] provides a ispector-executor approach to load balacig for irregular applicatios. Here the data ad the associated computatio balace is evaluated at rutime before the start of the first iteratio to rebalace. The proposed strategy is more focused towards iterative computatioal sciece applicatios, where computatioal tasks ted to be persistet. Dyamic load balacig algorithms for iterative applicatios ca be broadly classified as cetralized, distributed ad hierarchical. Cetralized strategies [7, 29] ted to yield good load balace but exhibit poor scalability. Alteratively, several distributed algorithms have bee proposed i which processors autoomously make load balacig decisios based o localized workload iformatio. Popular earest eighbor algorithms are dimesio-exchage [34] ad the diffusio methods. Dimesio-exchage method is performed i a iterative fashio ad is described i terms of a hypercube architecture. A processor performs load balacig with its eighbor i each dimesio of the hypercube. Diffusio based load balacig algorithms were first proposed by Cybeko [9] ad idepedetly by Boillat [4]. This algorithm suffers from slow covergece to the balaced state. Hu ad Blake [7] proposed a o-local method to determie the flow which is miimal i the l 2-orm but requires global commuicatio. The toke distributio problem was studied by Peleg ad Upfal [3] where the load is cosidered to be a toke. Several diffusive load balacig policies, like direct eighborhood, average eighborhood, have bee proposed i [8, 4, 9]. I [33], a seder-iitiated model is compared with receiver-iitiated i a asychroous settig. It also compares Gradiet Method [24], Hierarchical Method ad DEM (Dimesio exchage). The diffusio based load balacers are icremetal ad scale well with umber of processors. But, they ca be ivoked oly to improve load balace rather tha obtaiig global balace. If global balace is required, multiple iteratios might be required to coverge [5]. To overcome the disadvatages of cetralized ad distributed, hierarchical [35, 23, ] strategies have bee proposed. It is aother type of scheme which provides good performace ad scalig. I our proposed algorithm, global iformatio is spread usig a variat of gossip protocol []. Probabilistic gossipbased protocols have bee used as robust ad scalable methods for iformatio dissemiatio. Demers et al. use a gossip-based protocol to resolve icosistecies amog the Clearighouse database servers []. Birma et al. [2] employ gossip-based scheme for bi-modal multicast which they show to be reliable ad scalable. Apart from these, gossipbased protocols have bee adapted to implemet failure detectio, garbage collectio, aggregate computatio etc. 4. GRAPEVINE LOAD BALANCER Our distributed load balacig strategy, referred to as GrapevieLB, ca be coceptually thought of as havig two stages. ) Propagatio: Costructio of the local represetatio of the global state at each processor. 2) Trasfer: distributio based o the local represetatio. At the begiig of the load balacig step, the average load is calculated i parallel usig a efficiet tree based allreduce. This is followed by the propagatio stage, where the iformatio about the uderloaded processors i the system is spread to the overloaded processors. Oly the processor ID ad load of the uderloaded processors is propagated. A uderloaded processor starts the propagatio by selectig other processors radomly to sed iformatio. The receivig processors further spread the iformatio i a similar maer. Oce the overloaded processors have received the iformatio about the uderloaded processors, they autoomously make decisios about the trasfer of the work uits. Sice various processors do ot coordiate at this stage, the trasfer has to happe such that the probability that a uderloaded processor becomes overloaded is low. We propose a radomized algorithm that meets this goal. We elaborate further upo the above two stages i the followig sectios. 4. Iformatio propagatio To propagate the iformatio about the uderloaded processors i the system, GrapevieLB follows a protocol which is ispired by the epidemic algorithm [2] (also kow as the gossip protocol []). I our case, the goal is to spread the iformatio about the uderloaded processors such that every overloaded processor receives this iformatio with high probability. A uderloaded processor starts the ifectio by sedig its iformatio to a radomly chose subset of processors. The size of the subset is called faout, f. A ifected processor further spreads the ifectio by forwardig all the iformatio it has to aother set of radomly selected f processors. Here, each processor makes a idepedet radom selectio of peers to sed the iformatio. We show that the umber of rouds required for all processors to receive the iformatio with high probability is
4 Algorithm Iformed selectio at each processor P i P Iput: f - Faout L avg - Average load of the system. k - Target umber of rouds L i - of this processor : S Set of uderloaded processors 2: L of uderloaded processors 3: if (L i < L avg) the 4: S P i; L L i 5: Radomly sample { P,..., P f } P 6: Sed (S, L) to { P,..., P f } 7: ed if 8: for (roud = 2 k) do 9: if (received msg i previous roud) the : R P \ S Iformed selectio : Radomly sample { P,..., P f } R 2: Sed (S, L) to { P,..., P f } 3: ed if 4: ed for : whe (S ew, L ew) is received New message 2: S S S ew; L L L ew Merge iformatio O(log f ), where is the umber of processors. We propose two radomized strategies of peer selectio as described below. Note that although we discuss various strategies i terms of rouds for the sake of clarity, there is o explicit sychroizatio for rouds i our implemetatio. Naive Selectio: I this selectio strategy, each uderloaded processor idepedetly iitiates the propagatio by sedig its iformatio to a radomly selected set of f peers. A receivig processor updates its kowledge with the ew iformatio. It the radomly selects f processors, out of the total of processors, ad forwards its curret kowledge. This selectio may iclude other uderloaded processors. Iformed Selectio: This strategy is similar to the Naive strategy except that the selectio of peers to sed the iformatio is doe icorporatig the curret kowledge. Sice the curret kowledge icludes a partial list of uderloaded processors, the selectio process is biased to ot iclude these processors. This helps propagate iformatio to the overloaded processors i fewer umber of rouds. This strategy is depicted i Algorithm. 4.2 Probabilistic trasfer of load I our distributed scheme the decisio makig for trasfer of load is decetralized. Every processor eeds to make these decisios i isolatio give the iformatio from the propagatio stage. We propose two radomized schemes to trasfer load. Naive Trasfer: The simplest strategy to trasfer load is to select processors uiformly at radom from the list of uderloaded processors. A overloaded processor trasfers load util its load is below a specified threshold. The value of threshold idicates how much of a imbalace is acceptable. As oe would expect, this radom selectio results i overloadig processors whose load is closer to the average. This is illustrated i Figure ad described i detail i Sectio 7.. Iformed Trasfer: A more iformed trasfer ca be made by radomly selectig uderloaded processors based Algorithm 2 Iformed trasfer at each processor P i P Iput: O - Set of objects i this processor S - Set of uderloaded processors T - Threshold to trasfer L i - of this processor L avg - Average load of the system : Compute p j P j S Usig eq. 2 2: Compute F j = k<j p k Usig eq. 3 3: while (L i > (T L avg)) do 4: Select object O i O 5: Radomly sample X S usig F Usig eq. 4 6: if (L X + load(o i) < L avg) the 7: L X = L X + load(o i) 8: L i = L i load(o i) 9: O O \ O i : ed if : ed while o their iitial load. We achieve this by assigig to each processor a probability that is iversely proportioal to its load i the followig maer: p i = Z ( Li Z = N L avg ( ) Li L avg ) (2a) (2b) Here p i is the probability assiged to the ith processor, L i its load, L avg is the average load of the system ad Z is a ormalizatio costat. To select processors accordig to this distributio we use the iversio method for geeratig samples from a probability distributio. More formally if p(x) is a probability desity fuctio, the the cumulative distributio fuctio F (y) is defied as: F (y) = p(x < y) = y p(x)dx (3) Give a uiformly distributed radom sample r s [, ], a sample from the target distributio ca be computed by: y s = F (r s) (4) Usig the above, we radomly select the processors accordig to p i for trasferrig load. This is summarized i Algorithm 2. Figure illustrates the results. 4.3 Partial Propagatio A iterestig questio to ask is what happes if the overloaded processors have icomplete iformatio. This may happe with high probability if the propagatio stage is termiated earlier tha log rouds. We hypothesize that to obtai good load balace, iformatio about all the uderloaded processors is ot ecessary. A overloaded processor ca have a partial set of uderloaded processors ad still achieve good balace. We empirically cofirm our hypothesis by a set of experimets i Sectio Grapevie+ Eve though the scheme where every processor makes autoomous decisio for radomized trasfer of work is less
5 Naive Trasfer Probability Requests Iformed Trasfer Probability Requests (a) of (b) Probability Distributio (c) Trasfers Received (d) After LB Figure : (a) Iitial load of the uderloaded processors, (b) Probabilities assiged to each of the processors, (c) Work uits trasferred to each uderloaded processor, (d) Fial load of the uderloaded processors after trasfer. likely to cause uderloaded processors to become overloaded, this may still happe. To guaratee that oe of the uderloaded processors get overloaded after the trasfer, we propose a improvemet over the origial GrapevieLB strategy. I the improved scheme, referred to as Grapevie+LB, we employ a egative-ackowledgemet based mechaism to allow a uderloaded processor to reject a trasfer of work uit. For every potetial work uit trasfer, the seder iitially seds a message to the receiver which cotais details about the load of the work uit. The receiver, depedig o the curret load, chooses to either accept or reject. If acceptig the work uit makes the receiver overloaded, the it rejects with a Nack (egative-ackowledgemet). A seder o receivig a Nack will try to fid aother processor from the list of uderloaded processors. This trial is carried out for a limited umber of times after which the processor gives up. This scheme will esure that o uderloaded processor gets overloaded. Although this requires exchagig additioal messages, the cost is ot sigificat as the commuicatio is overlapped with the decisio makig process. 5. ANALYSIS OF THE ALGORITHM This sectio presets a aalysis of the iformatio propagatio algorithm. We cosider a system of processors ad, for simplicity, assume that the processors commuicate i sychroous rouds with a faout f. Note that i practice the commuicatio is asychroous (Sectio 6). We show that the expected umber of rouds required to propagate iformatio to all the processors i the system with high probability is O(log f ). Although we aalyze the case of sigle seder, the results are same for multiple seders sice they commuicate cocurretly ad idepedetly. I roud r =, oe processor iitiates the iformatio propagatio by sedig out f messages. I all successive rouds, each processor that received a message i the previous roud seds out f messages. We are iterested i the probability, p s, that ay processor P i received the message by the ed of roud s. We ca compute it by p s = q s, where q s is the probability that the processor P i did ot receive ay message by the ed of roud s. Probability that a processor P i did ot receive a message set by some other processor is ( ) ( ),. Further, the umber of messages set out i roud r is f r, sice the fa-out is f. Clearly, ( q = ) f (5) Therefore, the probability that P i did ot receive ay message i ay of the r {,..., s} rouds is s ( q s = ) f r ( = ) (f+f 2 +f 3 + +f s ) = r= ( ) f f s f ( ) γf s, Where γ = f f Here f s f s, f s. Takig log of both sides ( log q s γf s log ) γf s ( ) γf s q s exp Approximatig by the first two terms of the Taylor expasio of e x q s γf s Sice we wat to esure that the probability that a processor P i did ot receive ay message i s rouds is very low i.e. q s, substitutig this i the above yields γf s As q s s log f log log γ ( ) f s log f log f f = O(log f ) Our simulatio results show i figure 3 cocur with the above aalysis. It is evidet that icreasig the fa-out results i sigificat reductio of the umber of rouds required to propagate the iformatio.
6 6. IMPLEMENTATION We provide a implemetatio of the proposed algorithm as a load balacig strategy i Charm++. Charm++ is a parallel programmig model which has message drive parallel objects, chares, which ca be migrated from oe processor to aother. Chares are basic uits of parallel computatio i Charm++, which are mapped oto processors iitially usig a default mappig or ay custom mappig. --withig Charm++ load balacig framework supports istrumetig load iformatio of work uits from the recet past ad usig it as a guidelie for the ear future. The key advatage of this approach is that it is applicatio idepedet, ad has bee show to be effective for a large class of applicatios, such as NAMD [27] ad ChaNGa [8]. Charm++ has a user-friedly iterface for obtaiig dyamic measuremets about chares. The load balacers, which are pluggable modules i Charm++, ca use this istrumeted load iformatio to make the load balacig decisios. Based o these decisios Charm++ RTS migrates the chares. Sice the Charm++ RTS stores iformatio about chares ad processors i a distributed database, it is compatible with GrapevieLB s implemetatio requiremets. Although we have described the GrapevieLB algorithm i terms of rouds, a implemetatio usig barriers to eforce the rouds will icur cosiderable overhead. Therefore, we take a asychroous approach for our implemetatio. But such a approach poses the challege of limitig the umber of messages i the system. We overcome this by usig a T T L (Time To Live) based mechaism which limits the circulatio of iformatio forever. It is implemeted as a couter embedded i the messages beig propagated. The first message iitiated by a uderloaded processor is iitialized with the T T L of desired umber of rouds before beig set. A receivig processor icorporates the iformatio ad seds out a ew message with updated iformatio ad decremeted T T L. A message with T T L = is ot forwarded ad is cosidered expired. The key challege that remais is to detect quiescece, i.e. whe all the messages have expired. To this ed, we use a distributed termiatio detectio algorithm [26]. 7. EVALUATION We evaluate various stages of GrapevieLB with simulatios usig real data ad compare it with alterative strategies usig real world applicatios. 7. Evaluatio usig Simulatio We first preset results of simulatio of GrapevieLB strategy usig real data o a sigle processor. This simulatio allows us to demostrate the effect of various choices made i differet stages of the algorithm. For the simulatios, the system model is a set of 892 processors, iitialized with load from a real ru of a adaptive mesh refiemet applicatio with same umber of cores o IBM BG/Q. This applicatio was decomposed ito 253, 45 work uits. Figure 2 shows the load distributio for this applicatio whe the load balacer was ivoked. The average load of the system is 35, the maximum load is 66, therefore I, metric for imbalace from Equatio, is.88. Note that the value of I idicates perfect balace i the system. Amog the 892 processors, 495 are overloaded ad 497 are either uderloaded or have their Processors Figure 2: distributio for a ru of AMR used i simulatio. Couts of processors for various loads are depicted. Rouds f=2 4 f=3 f= System Size () Figure 3: Expected umber of rouds take to spread iformatio from oe source to 99% of the overloaded processors for differet system sizes ad faouts. load close to average. We perform a step-by-step aalysis of all the stages of the proposed algorithm based o this system model. It is to be oted that we have simulated sychroous rouds. The experimets were ru 5 times ad we report the results as mea alog with its stadard deviatio. Number of Rouds ad Faout: Figure 3 illustrates the depedece of expected umber of rouds required to spread iformatio o the system size. Here we cosider oly oe source iitiatig the propagatio ad report whe 99% of processors have received the iformatio. As the system size () icreases, the expected umber of rouds icrease logarithmically, O(log ), for a fixed faout. This is i accordace with our aalysis i Sectio 5. Note that the umber of rouds decreases with icrease i the faout used for the iformatio propagatio. A system size of 6K, faout of 2, requires 7 rouds to propagate iformatio to 99% processors whereas, faout of 4, takes 8 rouds. Naive vs Iformed Propagatio: Figure 4 compares the expected umber of rouds take to propagate iformatio usig Naive ad Iformed propagatio schemes. Although, the expected umber of rouds for both the schemes is o the order of O(log ), the Iformed scheme takes oe less roud to propagate the iformatio. This directly results i the reductio of the umber of messages as most of the messages are set i the later rouds. We ca also choose to vary the faout adaptively to reduce
7 Rouds System Size () Naive Iformed Max Max Imbalace Uderloaded Processor Ifo Imbalace Figure 4: Expected umber of rouds take to spread iformatio from oe source to 99% of the overloaded processors usig Naive ad Iformed schemes for differet system sizes. Here f = 2 ad 5% of the system size is uderloaded Figure 5: Evaluatio of load balacer with partial iformatio. Max load(left) ad Imbalace(right) decrease as more iformatio about uderloaded processors is available. It is evidet that complete iformatio is ot ecessary to obtai good performace. the umber of rouds required, while ot icreasig the umber of messages sigificatly. Istead of havig a fixed faout, we icrease the faout i the later stages. This is based o the observatio that messages i the iitial stages do ot carry a lot of iformatio. We evaluated this for a system of 496 processors where 5% were overloaded. Iformatio propagatio without the adaptive variatio requires 3 rouds with a total of 796 messages. While a adaptive faout strategy, where we use a faout of 2 iitially ad icrease the faout to 3 beyod 5 rouds ad further icrease to 4 beyod 7 rouds, helps reduce the umber of rouds to with a total of 864 messages. Naive vs Iformed Trasfer: We compare the performace of the two radomized strategies for trasfer give i Sectio 4. Figure shows the Naive scheme for the trasfer of load where a uderloaded processor is selected uiformly at radom. Here we also show the probability distributio of the uderloaded processors for the Iformed trasfer strategy usig the equatio 2 ad the trasfer of load which follows this distributio which are show i Figure. It shows the iitial load distributio of the uderloaded processors, probability assiged to each processor (uiform distributio), umber of trasfers based o the probability distributio ad the fial load of the uderloaded processors. It ca be see that the maximum load of the iitially uderloaded processors is 44 while the average is 35. Compariso with Figure clearly shows that the fial distributio of load is much more reasoable. Further, the maximum load of the uderloaded processors is 38 while the system average is 35. Evaluatio of a Pathological Case: We evaluate the behavior of the proposed algorithm uder the pathological case where just oe out of 892 processors is sigificatly overloaded (I is 6.8). Aalysis i Sectio 5 shows that q s decreases rapidly with rouds for a particular source. Sice all uderloaded processors will iitiate iformatio propagatio, this sceario should t be ay worse i expectatio. We experimetally verify this ad fid that for a faout value of 2 ad usig the Naive strategy for iformatio propagatio, it takes a maximum of 4 rouds to propagate the iformatio which is similar to the case where may processors are overloaded. Oce the iformatio is available at the overloaded processor, it radomly trasfers the work uits, reducig the I from 6.8 to.. Evaluatio of Quality of Balacig: To aswer the questio posed i the earlier sectio as to what happes if the overloaded processors have icomplete iformatio, we simulate this sceario by providig iformatio about oly a partial subset of uderloaded processors to the overloaded processors. The subset of uderloaded processors for each processor is selected uiformly at radom from the set of uderloaded processors ad the probabilistic trasfer of load is the carried out based o this partial iformatio. The quality is evaluated based o the metric I give by equatio. Figure 5 shows the expected maximum load of the system alog with stadard deviatio, σ ad the value of I metric. It ca be see that o oe had havig less iformatio, 5 uderloaded processors, yields cosiderable improvemet of load balace although ot the optimal possible. O the other had, havig complete iformatio is also ot ecessary to obtai good load balace. Therefore, this gives us a opportuity to trade-off betwee the overhead icurred ad load balace achieved. Evaluatio of Iformatio Propagatio: Based o the earlier experimet, it is evidet that complete iformatio about the uderloaded processors is ot required for good load balace. Therefore, we evaluate the expected umber of rouds take to propagate partial iformatio about the uderloaded processors to all the overloaded processors. Figure 6 shows the percetage of overloaded processors that received the iformatio as the rouds progress for a faout of 2. The x-axis is the umber of rouds ad the y- axis is the percetage of overloaded processors who received the iformatio. We plot the umber of rouds required to propagate iformatio about 2, 4, 248, 497 uderloaded processors to all the overloaded processors. I the case of propagatig iformatio about at least 2 uderloaded processors i the system, % of the overloaded processors receive iformatio about at least 2 uderloaded processors i 2 rouds ad 99.8% received i 9 rouds. It took 8 rouds to propagate iformatio about all the uderloaded processors i the system to all the overloaded processors. This clearly idicates that if we require oly partial iformatio, the total umber of rouds ca be reduced which will result i reductio of the load balacig cost.
8 % Processors Rouds Figure 6: Percetage of processors havig various amouts of partial iformatio as rouds progress. There are a total of 496 uderloaded processors. 99% receive iformatio about 4 processors by 8th roud while it takes 2 rouds for all the 496 uderloaded processors. From the above experimets, it is evidet that good load balace could be attaied with partial iformatio. This is particularly useful as propagatig partial iformatio takes fewer umber of rouds ad icurs lesser overhead. We utilize this observatio to choose a value of T T L much lower tha log for compariso with other strategies o real applicatios. 7.2 Evaluatio usig Applicatios We evaluate our GrapevieLB load balacig strategy o two applicatios, LeaMD ad adaptive mesh refiemet (AMR), by comparig agaist various load balacig strategies. We use GrapevieLB with a fixed set of cofiguratios, {f = 2, T T L =.4 log 2, Iformed Propagatio, Iformed Trasfer }, ad focus o comparig with other load balacig strategies. Results preseted here are obtaied from experimets ru o IBM BG/Q Mira. Mira is a 49, 52 ode Blue Gee/Q istallatio at the ALCF. Each ode cosists of 6 64-bit PowerPC A2 cores ru at.6ghz. The itercoect i this system is a 5D torus. I the followig sectios, we first provide details about the applicatios ad the load balacers ad the preset our evaluatio results Applicatios Adaptive Mesh Refiemet: AMR is a efficiet techique used to perform simulatios o very large meshes which would otherwise be difficult to simulate eve o moder-day supercomputers. This applicatio simulates a popular yet simple partial differetial equatio called Advectio. It uses a first-order upwid method i 2D space for solvig the advectio equatio. The simulatio begis o a coarse-graied structured grid of uiform size. As the simulatio progresses, idividual grids are either refied or coarseed. This leads to slowly-growig load imbalace which requires frequet load balacig to maitai high efficiecy of the system. This applicatio has bee implemeted usig the object-based decompositio approach i Charm++ [22]. LeaMD: It is a molecular dyamics simulatio program writte i Charm++, that simulates the behavior of atoms based o the Leard-Joes potetial. The computatios performed i this code are similar to the short-rage oboded force calculatio i NAMD [27], a applicatio that has wo the Gordo Bell award. The three-dimesioal simulatio space cosistig of atoms is divided ito cells. I each iteratio, force calculatios are doe for all pairs of atoms that are withi a specified cutoff distace. For a pair of cells, the force calculatio is assiged to a set of objects called the computes. After the force calculatio is performed by the computes, the cells update the acceleratio, velocity ad positio of the atoms withi their space. The load imbalace i LeaMD is primarily due to the variable umber of atoms i a cell. The load o computes is proportioal to the the umber of atoms i the cells which chages over time as the atoms move based o the force calculatio. We preset simulatio of LeaMD for a 2.8 millio atom system. The load imbalace is gradual therefore load balacig is performed ifrequetly Balacers We compare the performace of GrapevieLB agaist several other strategies icludig cetralized, distributed ad hierarchical strategies. The load balacig strategies are GreedyLB: A cetralized strategy that uses greedy heuristic to assig heaviest tasks oto least loaded processors iteratively. This strategy does ot take ito cosideratio the curret assigmet of tasks to processors. AmrLB: A cetralized strategy that does refiemet based load balacig takig ito accout the curret distributio of work uits. This is tued for the AMR applicatio [22]. HierchLB: A hierarchical strategy [35] i which processors are divided ito idepedet groups ad groups are orgaized i a hierarchical maer. At each level of the hierarchy, the root ode performs the load balacig for the processors i its sub-tree. This strategy ca use differet load balacig algorithms at differet levels. It is a optimized implemetatio that is used i strog scalig NAMD to more tha 2K cores. DiffusLB: A eighborhood averagig diffusio strategy [8, 33] where each processor seds iformatio to its eighbors i a domai ad load is exchaged based o this iformatio. A domai costitutes of a ode ad all its eighbors where the eighborhood is determied by physical topology. O receivig the load iformatio from all its eighbors, a ode will compute the average of the domai ad determies the amout of work uits to be trasfered to each of its eighbors. This is a two phase algorithm: i the first phase tokes are set ad i the secod phase actual movemet of work uits is performed. There are multiple iteratios of toke exchage ad termiatio is detected via quiescece [26]. We use the followig metrics to evaluate the performace of various load balacig strategies: ) Executio time per step for the applicatio, which idicates the quality of the load balacig strategy. 2) balacig overhead, which is the time take by a load balacig strategy. 3) Total applicatio time, which icludes the time for each iteratio as well as the time for load balacig strategy Evaluatio with AMR We preset a evaluatio of differet load balacig strategies o the AMR applicatio o BG/Q ragig from 496 to 372 cores. AMR requires frequet load balacig to ru efficietly because coarseig ad refiemet of the mesh itroduces dyamic load imbalace. Time per Iteratio: First we compare the executio
9 Time per Step (ms) No LB Diffus LB Amr LB Hierch LB GV LB GV+ LB LB Number of Cores 4K 8K 6K 32K 65K 3K No Hierc Amr Diff Gv Gv Number of Cores Table 3: Total applicatio time (i secods) for AMR o BG/Q. Proposed strategies Gv ad Gv+ perform the best across all scales. Figure 7: Compariso of time per step (excludig load balacig time) for various load balacig strategies for AMR o Mira (IBM BG/Q). GV+ achieves quality similar to other best performig strategies. Note that axes are log scale. LB Number of Cores 4K 8K 6K 32K 65K 3K Hierc Amr Diff Gv Gv Table 2: Average cost (i secods) per load balacig step of various strategies for AMR time per iteratio of the applicatio to evaluate the quality of the load balacers. This directly relates to I metric give i equatio because as I, the maximum load of the system approaches the average load, resultig i least time per iteratio. Figure 7 shows, o logarithmic scale, the time take per iteratio with various load balacig strategies. The base ru was made without ay load balacig ad is referred to as NoLB. It is evidet that with NoLB the efficiecy of the applicatio reduces as it is scaled to higher umber of cores. The Grapevie+LB load balacer (show as GV+ LB) reduces the iteratio time by 22% o 4K cores ad 5% o 3K cores. AmrLB ad HierchLB also show comparable performace for this metric. We see a icrease i gai because o larger umber of cores, the load imbalace becomes sigificat. This is because the umber of work uits per processor decreases ad the chace that a processor becomes overloaded icreases. DiffusLB also shows some improvemet but much less tha the aforemetioed oes o larger scale. For 3K, it reduces the time per step by 22% while others (AmrLB, HierchLB ad Grapevie+LB) reduce it by 5%. A iterestig thig to ote here is that, Grapevie+LB load balacer performs better tha GrapevieLB (show as GV LB) for core couts more tha 32K. This is due to the fact that Grapevie+LB esures that o uderloaded processor gets overloaded usig a Nack mechaism. From this it is evidet that the quality of load balace performed by Grapevie+LB is at-par with the quality of the cetralized ad hierarchical strategies. Overhead: Table 2 shows the overhead icurred by various load balacers i oe load balacig step for differet system sizes. The overhead(load balacig cost) icludes the time for fidig the ew assigmet of objects to processors ad the time for migratig the objects. The overhead icurred by AmrLB is 2. s for 4K cores ad icreases with the icrease i the system size to a maximum of 2.4 s for 3K cores. HierchLB icurs a overhead of 5.5 s for 8K cores ad thereafter the cost reduces to a miimum of.29 s for 3K cores. This is due to the fact that as the umber of processors icreases, the umber of sub groups also icrease resultig i a reductio of work uits per group. Hece, the time take for the root to carry out the load balacig strategy reduces. The distributed load balacig strategies, GrapevieLB ad DiffusLB, icur cosiderably less overhead i compariso to other strategies. Total Applicatio Time: The total applicatio time usig various strategies is give i Table 3. I this applicatio frequet load balacig is required. The overhead of the cetralized strategies dimiishes the beefit of load balacig. AmrLB does ot improve the total applicatio time because of the overhead of load balacig. This is true for the hierarchical strategy as well. The DiffusLB results i a reductio of the executio time by 28% for 6K cores ad 24.8% for 3K cores where as GrapevieLB gives a reductio of 35% ad 49.6% respectively. GrapevieLB provides a large performace gai by achievig a better load balace ad icurrig less overhead. It eables more frequet load balacig to improve the efficiecy. A future directio would be to use MetaBalacer [28] to choose the ideal load balacig period Evaluatio with LeaMD We evaluate LeaMD by executig a iteratios ad ivokig the load balacer first time at the th iteratio ad periodically every 3 iteratios there after. Executio time per iteratio: We compare the executio time per iteratio of the applicatio to evaluate the quality of the load balacers. For 4K to 6K cores, the cetralized, hierarchical ad GrapevieLB strategies improve the balace up to 42%. The diffusio-based strategy improves the balace oly by 35% at 8K cores ad there after it shows dimiishig gais. GrapevieLB o the other had performs at-par to the cetralized load balacer up to 32K. At 3K cores, it oly gives a improvemet of 25% i compariso to 36% give by cetralized scheme. This reductio is because the umber of tasks per processor decreases to 4 at 3K, causig refiemet-based load balacers to perform suboptimally. GrapevieLB is cosistetly better tha the DiffusLB because it has a represetatio of the global state of the system which helps it make better load balacig decisios.
10 Steps per Secod Performace of LeaMD o BlueGee/Q No LB Refie LB GV+ LB Nbor LB Hybrid LB Greedy LB GV LB LB Number of Cores 4K 8K 6K 32K 65K 3K No Hierc Grdy Diff Gv Gv Table 5: Total applicatio time (i secods) for LeaMD o BG/Q Number of processes Figure 8: Compariso of time per step (excludig load balacig time) for various load balacig strategies for LeaMD o Mira (IBM BG/Q). Note that axes are log scale. LB Number of Cores 4K 8K 6K 32K 65K 3K Hierc Grdy Diff Gv Gv Table 4: Average cost per load balacig step (i secods) of various strategies for LeaMD Overhead: Table 4 presets a compariso of overhead icurred by various strategies for a sigle load balacig step. The load balacig cost of the cetralized strategy is very high ad is o the order of tes of secods. The high overhead of GreedyLB is due to the overhead of statistics collectio, makig the decisio at the cetral locatio ad the migratio cost. The hierarchical strategy, HierchLB, icurs less overhead. It takes 3.7 s for 4K cores ad decreases to.26 s as the system size icreases to 3K. The overhead of DiffusLB is.8 s for 4K cores ad decreases thereafter. This is because the umber of work uits per core decreases as the umber of cores icrease. Fially, we observe that GrapevieLB has a overhead of.7 s for 4K cores ad decreases with icrease i system size to.3 s for 6K cores ad thereafter icreases to.8 s for 3K. The load balacig cost for GrapevieLB icludes the time for iformatio propagatio ad trasfer of work uits. At 4K cores the load balacig time is domiated by the trasfer of work uits. As the system size icreases, the work uits per processor decreases. This results i cost beig domiated by iformatio propagatio. Total Applicatio Time: Table 5 shows the total applicatio time for LeaMD. The cetralized strategy improves the total applicatio time but oly for core couts up to 6K. Beyod 6K cores, the overhead due to load balacig exceeds the gais ad results i icreasig the total applicatio time. DiffusLB icurs less overhead i compariso to the cetralized ad hierarchical strategies but it does ot show substatial gais because the quality of load balace is ot good. At 32K cores, it gives a reductio of 2% i total executio time while GrapevieLB gives 34% ad HierchLB gives 33%. HierchLB icurs less overhead i compariso to the cetralized strategies. It reduces the total execu- tio time by 37% for 8K cores while GrapevieLB reduces it by 42%. GrapevieLB cosistetly gives better performace tha other load balacig strategies. Grapevie+LB gives the maximum performace beefit by reducig the total applicatio time by 2% for 3K, 4% for 6K cores, aroud 42% for 4K ad 8K cores. Thus, GrapevieLB ad Grapevie+LB provide a improvemet i performace by achievig a high quality load balace with sigificatly less overhead. 8. CONCLUSION We have preseted GrapevieLB, a ovel algorithm for distributed load balacig. It icludes a light weight iformatio propagatio stage based o gossip protocol to obtai partial iformatio about the global state of the system. Exploitig this iformatio, GrapevieLB probabilistically trasfers work uits to obtai high quality load distributio. We have demostrated performace gais of GrapevieLB by comparig agaist various cetralized, distributed ad hierarchical load balacig strategies for molecular dyamics simulatio ad adaptive mesh refiemet. GrapevieLB is show to match the quality of cetralized strategies, i terms of the time per iteratio, while avoidig associated bottleecks. Our experimets demostrate that it sigificatly reduces the total applicatio time i compariso to other load balacig strategies as it achieves good load distributio while icurrig less overhead. Ackowledgmet The authors would like to thak Phil Miller, Joatha Lifflader ad Nikhil Jai for their valuable help i proofreadig. This research was supported i part by the US Departmet of Eergy uder grat DOE DE-SC845 ad by NSF ITR-HECURA This research also used resources of the Argoe Leadership Computig Facility at Argoe Natioal Laboratory, which is supported by the Office of Sciece of the U.S. Departmet of Eergy uder cotract DE-AC2-6CH357. Experimets for this work were performed o Mira ad esta, IBM Blue Gee/Q istallatios at Argoe Natioal Laboratory. The authors would like to ackowledge PEACEdStatio ad PARTS projects for the machie allocatios provided by them. 9. REFERENCES [] I. Ahmad ad A. Ghafoor. A semi distributed task allocatio strategy for large hypercube supercomputers. I Coferece o Supercomputig, 99.
11 [2] K. Birma, M. Hayde, O. Ozkasap, Z. Xiao, M. Budiu, ad Y. Misky. Bimodal multicast. ACM Trasactios o Computer Systems (TOCS), 999. [3] R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leiserso, K. H. Radall, ad Y. Zhou. Cilk: A Efficiet Multithreaded Rutime System. I PPoPP, 995. [4] J. E. Boillat. balacig ad poisso equatio i a graph. Cocurrecy: Practice ad Experiece, 2(4):289 33, 99. [5] U. Catalyurek, E. Boma, K. Devie, D. Bozdag, R. Heaphy, ad L. Riese. Hypergraph-based dyamic load balacig for adaptive scietific computatios. I Proc. of 2st Iteratioal Parallel ad Distributed Processig Symposium (IPDPS 7), pages. IEEE, 27. Best Algorithms Paper Award. [6] C. Chevalier, F. Pellegrii, I. Futurs, ad U. B. I. Improvemet of the efficiecy of geetic algorithms for scalable parallel graph partitioig i a multi-level framework. I I Proceedigs of Euro-Par 26, LNCS, pages , 26. [7] Y.-C. Chow ad W. H. Kohler. Models for dyamic load balacig i homogeeous multiple processor systems. I IEEE Trasactios o Computers, 982. [8] A. Corradi, L. Leoardi, ad F. Zamboelli. Diffusive load balacig policies for dyamic applicatios. I IEEE Cocurrecy, pages 7():22 3, 999. [9] G. Cybeko. Dyamic load balacig for distributed memory multiprocessors. Joural of parallel ad distributed computig, 7(2):279 3, 989. [] A. Demers, D. Greee, C. Hauser, W. Irish, J. Larso, S. Sheker, H. Sturgis, D. Swiehart, ad D. Terry. Epidemic algorithms for replicated database maiteace. I ACM Symposium o Priciples of distributed computig, 987. [] J. Dia, D. B. Larkis, P. Sadayappa, S. Krishamoorthy, ad J. Nieplocha. Scalable work stealig. I Coferece o High Performace Computig Networkig, Storage ad Aalysis, 29. [2] George Karypis ad Vipi Kumar. A coarse-grai parallel formulatio of multilevel k-way graph partitioig algorithm. I Proc. of the 8th SIAM coferece o Parallel Processig for Scietific Computig, 997. [3] George Karypis ad Vipi Kumar. Multilevel k-way Partitioig Scheme for Irregular Graphs. Joural of Parallel ad Distributed Computig, 48:96 29, 998. [4] A. Ha c ad X. Ji. Dyamic load balacig i distributed system usig a decetralized algorithm. I Itl. Cof. o Distributed Computig Systems, 987. [5] B. Hedrickso ad K. Devie. Dyamic load balacig i computatioal mechaics. Computer Methods i Applied Mechaics ad Egieerig, 84(2):485 5, 2. [6] B. Hedrickso ad R. Lelad. The Chaco user s guide. Techical Report SAND , Sadia Natioal Laboratories, Albuquerque, NM, Oct [7] Y. Hu ad R. Blake. A optimal dyamic load balacig algorithm. Techical report, Daresbury Laboratory, 995. [8] P. Jetley, F. Gioachi, C. Medes, L. V. Kale, ad T. R. Qui. Massively parallel cosmological simulatios with ChaNGa. I IPDPS, 28. [9] L. V. Kalé. Comparig the performace of two dyamic load distributio methods. I Proceedigs of the 988 Iteratioal Coferece o Parallel Processig, pages 8, St. Charles, IL, August 988. [2] L. V. Kalé. The virtualizatio model of parallel programmig : Rutime optimizatios ad the state of art. I LACSI 22, Albuquerque, October 22. [2] W. Kermack ad A. McKedrick. Cotributios to the mathematical theory of epidemics. ii. the problem of edemicity. Proceedigs of the Royal society of Lodo. Series A, 38(834):55 83, 932. [22] A. Lager, J. Lifflader, P. Miller, K.-C. Pa, L. V. Kale, ad P. Ricker. Scalable Algorithms for Distributed-Memory Adaptive Mesh Refiemet. I SBAC-PAD 22, New York, USA, October 22. [23] J. Lifflader, S. Krishamoorthy, ad L. V. Kale. Work stealig ad persistece-based load balacers for iterative overdecomposed applicatios. I HPDC, 22. [24] F. C. H. Li ad R. M. Keller. The gradiet model load balacig method. Software Egieerig, IEEE Trasactios o, ():32 38, 987. [25] Y.-J. Li ad V. Kumar. Ad-parallel executio of logic programs o a shared-memory multiprocessor. J. Log. Program., (/2/3&4):55 78, 99. [26] F. Matter. Algorithms for distributed termiatio detectio. Distributed computig, 2(3):6 75, 987. [27] C. Mei ad L. V. K. et al. Eablig ad scalig biomolecular simulatios of millio atoms o petascale machies with a multicore-optimized message-drive rutime. I Proceedigs of the 2 ACM/IEEE coferece o Supercomputig. [28] H. Meo, N. Jai, G. Zheg, ad L. V. Kalé. Automated load balacig ivocatio based o applicatio characteristics. I IEEE Cluster, 22. [29] L. M. Ni ad K. Hwag. Optimal load balacig i a multiple processor system with may job classes. I IEEE Tras. o Software Eg., volume SE-, 985. [3] D. Peleg ad E. Upfal. The toke distributio problem. SIAM Joural o Computig, 8(2): , 989. [3] S. Sharma, R. Pousamy, B. Moo, Y. Hwag, R. Das, ad J. Saltz. Ru-time ad compile-time support for adaptive irregular problems. I Proceedigs of Supercomputig 994, Nov [32] Y. Su, G. Zheg, P. Jetley, ad L. V. Kale. A Adaptive Framework for Large-scale State Space Search. I IPDPS, 2. [33] M. H. Willebeek-LeMair ad A. P. Reeves. Strategies for dyamic load balacig o highly parallel computers. I IEEE Trasactios o Parallel ad Distributed Systems, September 993. [34] C. Xu, F. C. M. Lau, ad R. Diekma. Decetralized remappig of data parallel applicatios i distributed memory multiprocessors. Cocurrecy - Practice ad Experiece, 9(2):35 376, 997. [35] G. Zheg, A. Bhatele, E. Meeses, ad L. V. Kale. Periodic Hierarchical Balacig for Large Supercomputers. IJHPCA, March 2.
Modified Line Search Method for Global Optimization
Modified Lie Search Method for Global Optimizatio Cria Grosa ad Ajith Abraham Ceter of Excellece for Quatifiable Quality of Service Norwegia Uiversity of Sciece ad Techology Trodheim, Norway {cria, ajith}@q2s.tu.o
Domain 1: Designing a SQL Server Instance and a Database Solution
Maual SQL Server 2008 Desig, Optimize ad Maitai (70-450) 1-800-418-6789 Domai 1: Desigig a SQL Server Istace ad a Database Solutio Desigig for CPU, Memory ad Storage Capacity Requiremets Whe desigig a
CHAPTER 3 THE TIME VALUE OF MONEY
CHAPTER 3 THE TIME VALUE OF MONEY OVERVIEW A dollar i the had today is worth more tha a dollar to be received i the future because, if you had it ow, you could ivest that dollar ad ear iterest. Of all
Hypothesis testing. Null and alternative hypotheses
Hypothesis testig Aother importat use of samplig distributios is to test hypotheses about populatio parameters, e.g. mea, proportio, regressio coefficiets, etc. For example, it is possible to stipulate
Output Analysis (2, Chapters 10 &11 Law)
B. Maddah ENMG 6 Simulatio 05/0/07 Output Aalysis (, Chapters 10 &11 Law) Comparig alterative system cofiguratio Sice the output of a simulatio is radom, the comparig differet systems via simulatio should
Analyzing Longitudinal Data from Complex Surveys Using SUDAAN
Aalyzig Logitudial Data from Complex Surveys Usig SUDAAN Darryl Creel Statistics ad Epidemiology, RTI Iteratioal, 312 Trotter Farm Drive, Rockville, MD, 20850 Abstract SUDAAN: Software for the Statistical
INVESTMENT PERFORMANCE COUNCIL (IPC)
INVESTMENT PEFOMANCE COUNCIL (IPC) INVITATION TO COMMENT: Global Ivestmet Performace Stadards (GIPS ) Guidace Statemet o Calculatio Methodology The Associatio for Ivestmet Maagemet ad esearch (AIM) seeks
Department of Computer Science, University of Otago
Departmet of Computer Sciece, Uiversity of Otago Techical Report OUCS-2006-09 Permutatios Cotaiig May Patters Authors: M.H. Albert Departmet of Computer Sciece, Uiversity of Otago Micah Colema, Rya Fly
DAME - Microsoft Excel add-in for solving multicriteria decision problems with scenarios Radomir Perzina 1, Jaroslav Ramik 2
Itroductio DAME - Microsoft Excel add-i for solvig multicriteria decisio problems with scearios Radomir Perzia, Jaroslav Ramik 2 Abstract. The mai goal of every ecoomic aget is to make a good decisio,
COMPARISON OF THE EFFICIENCY OF S-CONTROL CHART AND EWMA-S 2 CONTROL CHART FOR THE CHANGES IN A PROCESS
COMPARISON OF THE EFFICIENCY OF S-CONTROL CHART AND EWMA-S CONTROL CHART FOR THE CHANGES IN A PROCESS Supraee Lisawadi Departmet of Mathematics ad Statistics, Faculty of Sciece ad Techoology, Thammasat
INVESTMENT PERFORMANCE COUNCIL (IPC) Guidance Statement on Calculation Methodology
Adoptio Date: 4 March 2004 Effective Date: 1 Jue 2004 Retroactive Applicatio: No Public Commet Period: Aug Nov 2002 INVESTMENT PERFORMANCE COUNCIL (IPC) Preface Guidace Statemet o Calculatio Methodology
LECTURE 13: Cross-validation
LECTURE 3: Cross-validatio Resampli methods Cross Validatio Bootstrap Bias ad variace estimatio with the Bootstrap Three-way data partitioi Itroductio to Patter Aalysis Ricardo Gutierrez-Osua Texas A&M
Determining the sample size
Determiig the sample size Oe of the most commo questios ay statisticia gets asked is How large a sample size do I eed? Researchers are ofte surprised to fid out that the aswer depeds o a umber of factors
Domain 1 - Describe Cisco VoIP Implementations
Maual ONT (642-8) 1-800-418-6789 Domai 1 - Describe Cisco VoIP Implemetatios Advatages of VoIP Over Traditioal Switches Voice over IP etworks have may advatages over traditioal circuit switched voice etworks.
Optimize your Network. In the Courier, Express and Parcel market ADDING CREDIBILITY
Optimize your Network I the Courier, Express ad Parcel market ADDING CREDIBILITY Meetig today s challeges ad tomorrow s demads Aswers to your key etwork challeges ORTEC kows the highly competitive Courier,
Vladimir N. Burkov, Dmitri A. Novikov MODELS AND METHODS OF MULTIPROJECTS MANAGEMENT
Keywords: project maagemet, resource allocatio, etwork plaig Vladimir N Burkov, Dmitri A Novikov MODELS AND METHODS OF MULTIPROJECTS MANAGEMENT The paper deals with the problems of resource allocatio betwee
Taking DCOP to the Real World: Efficient Complete Solutions for Distributed Multi-Event Scheduling
Taig DCOP to the Real World: Efficiet Complete Solutios for Distributed Multi-Evet Schedulig Rajiv T. Maheswara, Milid Tambe, Emma Bowrig, Joatha P. Pearce, ad Pradeep araatham Uiversity of Souther Califoria
Convention Paper 6764
Audio Egieerig Society Covetio Paper 6764 Preseted at the 10th Covetio 006 May 0 3 Paris, Frace This covetio paper has bee reproduced from the author's advace mauscript, without editig, correctios, or
Case Study. Normal and t Distributions. Density Plot. Normal Distributions
Case Study Normal ad t Distributios Bret Halo ad Bret Larget Departmet of Statistics Uiversity of Wiscosi Madiso October 11 13, 2011 Case Study Body temperature varies withi idividuals over time (it ca
Lesson 17 Pearson s Correlation Coefficient
Outlie Measures of Relatioships Pearso s Correlatio Coefficiet (r) -types of data -scatter plots -measure of directio -measure of stregth Computatio -covariatio of X ad Y -uique variatio i X ad Y -measurig
(VCP-310) 1-800-418-6789
Maual VMware Lesso 1: Uderstadig the VMware Product Lie I this lesso, you will first lear what virtualizatio is. Next, you ll explore the products offered by VMware that provide virtualizatio services.
LEASE-PURCHASE DECISION
Public Procuremet Practice STANDARD The decisio to lease or purchase should be cosidered o a case-by case evaluatio of comparative costs ad other factors. 1 Procuremet should coduct a cost/ beefit aalysis
In nite Sequences. Dr. Philippe B. Laval Kennesaw State University. October 9, 2008
I ite Sequeces Dr. Philippe B. Laval Keesaw State Uiversity October 9, 2008 Abstract This had out is a itroductio to i ite sequeces. mai de itios ad presets some elemetary results. It gives the I ite Sequeces
Reliability Analysis in HPC clusters
Reliability Aalysis i HPC clusters Narasimha Raju, Gottumukkala, Yuda Liu, Chokchai Box Leagsuksu 1, Raja Nassar, Stephe Scott 2 College of Egieerig & Sciece, Louisiaa ech Uiversity Oak Ridge Natioal Lab
Soving Recurrence Relations
Sovig Recurrece Relatios Part 1. Homogeeous liear 2d degree relatios with costat coefficiets. Cosider the recurrece relatio ( ) T () + at ( 1) + bt ( 2) = 0 This is called a homogeeous liear 2d degree
Hypergeometric Distributions
7.4 Hypergeometric Distributios Whe choosig the startig lie-up for a game, a coach obviously has to choose a differet player for each positio. Similarly, whe a uio elects delegates for a covetio or you
Project Deliverables. CS 361, Lecture 28. Outline. Project Deliverables. Administrative. Project Comments
Project Deliverables CS 361, Lecture 28 Jared Saia Uiversity of New Mexico Each Group should tur i oe group project cosistig of: About 6-12 pages of text (ca be loger with appedix) 6-12 figures (please
Recovery time guaranteed heuristic routing for improving computation complexity in survivable WDM networks
Computer Commuicatios 30 (2007) 1331 1336 wwwelseviercom/locate/comcom Recovery time guarateed heuristic routig for improvig computatio complexity i survivable WDM etworks Lei Guo * College of Iformatio
How to read A Mutual Fund shareholder report
Ivestor BulletI How to read A Mutual Fud shareholder report The SEC s Office of Ivestor Educatio ad Advocacy is issuig this Ivestor Bulleti to educate idividual ivestors about mutual fud shareholder reports.
Systems Design Project: Indoor Location of Wireless Devices
Systems Desig Project: Idoor Locatio of Wireless Devices Prepared By: Bria Murphy Seior Systems Sciece ad Egieerig Washigto Uiversity i St. Louis Phoe: (805) 698-5295 Email: [email protected] Supervised
1 Computing the Standard Deviation of Sample Means
Computig the Stadard Deviatio of Sample Meas Quality cotrol charts are based o sample meas ot o idividual values withi a sample. A sample is a group of items, which are cosidered all together for our aalysis.
Designing Incentives for Online Question and Answer Forums
Desigig Icetives for Olie Questio ad Aswer Forums Shaili Jai School of Egieerig ad Applied Scieces Harvard Uiversity Cambridge, MA 0238 USA [email protected] Yilig Che School of Egieerig ad Applied
On the Capacity of Hybrid Wireless Networks
O the Capacity of Hybrid ireless Networks Beyua Liu,ZheLiu +,DoTowsley Departmet of Computer Sciece Uiversity of Massachusetts Amherst, MA 0002 + IBM T.J. atso Research Ceter P.O. Box 704 Yorktow Heights,
Automatic Tuning for FOREX Trading System Using Fuzzy Time Series
utomatic Tuig for FOREX Tradig System Usig Fuzzy Time Series Kraimo Maeesilp ad Pitihate Soorasa bstract Efficiecy of the automatic currecy tradig system is time depedet due to usig fixed parameters which
Effective Techniques for Message Reduction and Load Balancing in Distributed Graph Computation
Effective Techiques for Message Reductio ad Load Balacig i Distributed Graph Computatio ABSTRACT Da Ya, James Cheg, Yi Lu Dept. of Computer Sciece ad Egieerig The Chiese Uiversity of Hog Kog {yada, jcheg,
Spam Detection. A Bayesian approach to filtering spam
Spam Detectio A Bayesia approach to filterig spam Kual Mehrotra Shailedra Watave Abstract The ever icreasig meace of spam is brigig dow productivity. More tha 70% of the email messages are spam, ad it
5: Introduction to Estimation
5: Itroductio to Estimatio Cotets Acroyms ad symbols... 1 Statistical iferece... Estimatig µ with cofidece... 3 Samplig distributio of the mea... 3 Cofidece Iterval for μ whe σ is kow before had... 4 Sample
NEW HIGH PERFORMANCE COMPUTATIONAL METHODS FOR MORTGAGES AND ANNUITIES. Yuri Shestopaloff,
NEW HIGH PERFORMNCE COMPUTTIONL METHODS FOR MORTGGES ND NNUITIES Yuri Shestopaloff, Geerally, mortgage ad auity equatios do ot have aalytical solutios for ukow iterest rate, which has to be foud usig umerical
*The most important feature of MRP as compared with ordinary inventory control analysis is its time phasing feature.
Itegrated Productio ad Ivetory Cotrol System MRP ad MRP II Framework of Maufacturig System Ivetory cotrol, productio schedulig, capacity plaig ad fiacial ad busiess decisios i a productio system are iterrelated.
SECTION 1.5 : SUMMATION NOTATION + WORK WITH SEQUENCES
SECTION 1.5 : SUMMATION NOTATION + WORK WITH SEQUENCES Read Sectio 1.5 (pages 5 9) Overview I Sectio 1.5 we lear to work with summatio otatio ad formulas. We will also itroduce a brief overview of sequeces,
Estimating Probability Distributions by Observing Betting Practices
5th Iteratioal Symposium o Imprecise Probability: Theories ad Applicatios, Prague, Czech Republic, 007 Estimatig Probability Distributios by Observig Bettig Practices Dr C Lych Natioal Uiversity of Irelad,
PROCEEDINGS OF THE YEREVAN STATE UNIVERSITY AN ALTERNATIVE MODEL FOR BONUS-MALUS SYSTEM
PROCEEDINGS OF THE YEREVAN STATE UNIVERSITY Physical ad Mathematical Scieces 2015, 1, p. 15 19 M a t h e m a t i c s AN ALTERNATIVE MODEL FOR BONUS-MALUS SYSTEM A. G. GULYAN Chair of Actuarial Mathematics
CHAPTER 3 DIGITAL CODING OF SIGNALS
CHAPTER 3 DIGITAL CODING OF SIGNALS Computers are ofte used to automate the recordig of measuremets. The trasducers ad sigal coditioig circuits produce a voltage sigal that is proportioal to a quatity
Center, Spread, and Shape in Inference: Claims, Caveats, and Insights
Ceter, Spread, ad Shape i Iferece: Claims, Caveats, ad Isights Dr. Nacy Pfeig (Uiversity of Pittsburgh) AMATYC November 2008 Prelimiary Activities 1. I would like to produce a iterval estimate for the
Lecture 2: Karger s Min Cut Algorithm
priceto uiv. F 3 cos 5: Advaced Algorithm Desig Lecture : Karger s Mi Cut Algorithm Lecturer: Sajeev Arora Scribe:Sajeev Today s topic is simple but gorgeous: Karger s mi cut algorithm ad its extesio.
Normal Distribution.
Normal Distributio www.icrf.l Normal distributio I probability theory, the ormal or Gaussia distributio, is a cotiuous probability distributio that is ofte used as a first approimatio to describe realvalued
Chapter 6: Variance, the law of large numbers and the Monte-Carlo method
Chapter 6: Variace, the law of large umbers ad the Mote-Carlo method Expected value, variace, ad Chebyshev iequality. If X is a radom variable recall that the expected value of X, E[X] is the average value
Measures of Spread and Boxplots Discrete Math, Section 9.4
Measures of Spread ad Boxplots Discrete Math, Sectio 9.4 We start with a example: Example 1: Comparig Mea ad Media Compute the mea ad media of each data set: S 1 = {4, 6, 8, 10, 1, 14, 16} S = {4, 7, 9,
5 Boolean Decision Trees (February 11)
5 Boolea Decisio Trees (February 11) 5.1 Graph Coectivity Suppose we are give a udirected graph G, represeted as a boolea adjacecy matrix = (a ij ), where a ij = 1 if ad oly if vertices i ad j are coected
CS100: Introduction to Computer Science
Review: History of Computers CS100: Itroductio to Computer Sciece Maiframes Miicomputers Lecture 2: Data Storage -- Bits, their storage ad mai memory Persoal Computers & Workstatios Review: The Role of
Dynamic House Allocation
Dyamic House Allocatio Sujit Gujar 1 ad James Zou 2 ad David C. Parkes 3 Abstract. We study a dyamic variat o the house allocatio problem. Each aget ows a distict object (a house) ad is able to trade its
Capacity of Wireless Networks with Heterogeneous Traffic
Capacity of Wireless Networks with Heterogeeous Traffic Migyue Ji, Zheg Wag, Hamid R. Sadjadpour, J.J. Garcia-Lua-Aceves Departmet of Electrical Egieerig ad Computer Egieerig Uiversity of Califoria, Sata
Effective Techniques for Message Reduction and Load Balancing in Distributed Graph Computation
Effective Techiques for Message Reductio ad Load Balacig i Distributed Graph Computatio ABSTRACT Da Ya, James Cheg, Yi Lu Dept. of Computer Sciece ad Egieerig The Chiese Uiversity of Hog Kog {yada, jcheg,
A Combined Continuous/Binary Genetic Algorithm for Microstrip Antenna Design
A Combied Cotiuous/Biary Geetic Algorithm for Microstrip Atea Desig Rady L. Haupt The Pesylvaia State Uiversity Applied Research Laboratory P. O. Box 30 State College, PA 16804-0030 [email protected] Abstract:
Amendments to employer debt Regulations
March 2008 Pesios Legal Alert Amedmets to employer debt Regulatios The Govermet has at last issued Regulatios which will amed the law as to employer debts uder s75 Pesios Act 1995. The amedig Regulatios
The Stable Marriage Problem
The Stable Marriage Problem William Hut Lae Departmet of Computer Sciece ad Electrical Egieerig, West Virgiia Uiversity, Morgatow, WV [email protected] 1 Itroductio Imagie you are a matchmaker,
Pre-Suit Collection Strategies
Pre-Suit Collectio Strategies Writte by Charles PT Phoeix How to Decide Whether to Pursue Collectio Calculatig the Value of Collectio As with ay busiess litigatio, all factors associated with the process
Subject CT5 Contingencies Core Technical Syllabus
Subject CT5 Cotigecies Core Techical Syllabus for the 2015 exams 1 Jue 2014 Aim The aim of the Cotigecies subject is to provide a groudig i the mathematical techiques which ca be used to model ad value
I. Chi-squared Distributions
1 M 358K Supplemet to Chapter 23: CHI-SQUARED DISTRIBUTIONS, T-DISTRIBUTIONS, AND DEGREES OF FREEDOM To uderstad t-distributios, we first eed to look at aother family of distributios, the chi-squared distributios.
MTO-MTS Production Systems in Supply Chains
NSF GRANT #0092854 NSF PROGRAM NAME: MES/OR MTO-MTS Productio Systems i Supply Chais Philip M. Kamisky Uiversity of Califoria, Berkeley Our Kaya Uiversity of Califoria, Berkeley Abstract: Icreasig cost
The following example will help us understand The Sampling Distribution of the Mean. C1 C2 C3 C4 C5 50 miles 84 miles 38 miles 120 miles 48 miles
The followig eample will help us uderstad The Samplig Distributio of the Mea Review: The populatio is the etire collectio of all idividuals or objects of iterest The sample is the portio of the populatio
A Guide to the Pricing Conventions of SFE Interest Rate Products
A Guide to the Pricig Covetios of SFE Iterest Rate Products SFE 30 Day Iterbak Cash Rate Futures Physical 90 Day Bak Bills SFE 90 Day Bak Bill Futures SFE 90 Day Bak Bill Futures Tick Value Calculatios
Professional Networking
Professioal Networkig 1. Lear from people who ve bee where you are. Oe of your best resources for etworkig is alumi from your school. They ve take the classes you have take, they have bee o the job market
NATIONAL SENIOR CERTIFICATE GRADE 12
NATIONAL SENIOR CERTIFICATE GRADE MATHEMATICS P EXEMPLAR 04 MARKS: 50 TIME: 3 hours This questio paper cosists of 8 pages ad iformatio sheet. Please tur over Mathematics/P DBE/04 NSC Grade Eemplar INSTRUCTIONS
C.Yaashuwanth Department of Electrical and Electronics Engineering, Anna University Chennai, Chennai 600 025, India..
(IJCSIS) Iteratioal Joural of Computer Sciece ad Iformatio Security, A New Schedulig Algorithms for Real Time Tasks C.Yaashuwath Departmet of Electrical ad Electroics Egieerig, Aa Uiversity Cheai, Cheai
ODBC. Getting Started With Sage Timberline Office ODBC
ODBC Gettig Started With Sage Timberlie Office ODBC NOTICE This documet ad the Sage Timberlie Office software may be used oly i accordace with the accompayig Sage Timberlie Office Ed User Licese Agreemet.
Evaluation of Different Fitness Functions for the Evolutionary Testing of an Autonomous Parking System
Evaluatio of Differet Fitess Fuctios for the Evolutioary Testig of a Autoomous Parkig System Joachim Wegeer 1, Oliver Bühler 2 1 DaimlerChrysler AG, Research ad Techology, Alt-Moabit 96 a, D-1559 Berli,
The Forgotten Middle. research readiness results. Executive Summary
The Forgotte Middle Esurig that All Studets Are o Target for College ad Career Readiess before High School Executive Summary Today, college readiess also meas career readiess. While ot every high school
Sequences and Series
CHAPTER 9 Sequeces ad Series 9.. Covergece: Defiitio ad Examples Sequeces The purpose of this chapter is to itroduce a particular way of geeratig algorithms for fidig the values of fuctios defied by their
where: T = number of years of cash flow in investment's life n = the year in which the cash flow X n i = IRR = the internal rate of return
EVALUATING ALTERNATIVE CAPITAL INVESTMENT PROGRAMS By Ke D. Duft, Extesio Ecoomist I the March 98 issue of this publicatio we reviewed the procedure by which a capital ivestmet project was assessed. The
France caters to innovative companies and offers the best research tax credit in Europe
1/5 The Frech Govermet has three objectives : > improve Frace s fiscal competitiveess > cosolidate R&D activities > make Frace a attractive coutry for iovatio Tax icetives have become a key elemet of public
Incremental calculation of weighted mean and variance
Icremetal calculatio of weighted mea ad variace Toy Fich [email protected] [email protected] Uiversity of Cambridge Computig Service February 009 Abstract I these otes I eplai how to derive formulae for umerically
Quadrat Sampling in Population Ecology
Quadrat Samplig i Populatio Ecology Backgroud Estimatig the abudace of orgaisms. Ecology is ofte referred to as the "study of distributio ad abudace". This beig true, we would ofte like to kow how may
Predictive Modeling Data. in the ACT Electronic Student Record
Predictive Modelig Data i the ACT Electroic Studet Record overview Predictive Modelig Data Added to the ACT Electroic Studet Record With the release of studet records i September 2012, predictive modelig
University of California, Los Angeles Department of Statistics. Distributions related to the normal distribution
Uiversity of Califoria, Los Ageles Departmet of Statistics Statistics 100B Istructor: Nicolas Christou Three importat distributios: Distributios related to the ormal distributio Chi-square (χ ) distributio.
Non-life insurance mathematics. Nils F. Haavardsson, University of Oslo and DNB Skadeforsikring
No-life isurace mathematics Nils F. Haavardsso, Uiversity of Oslo ad DNB Skadeforsikrig Mai issues so far Why does isurace work? How is risk premium defied ad why is it importat? How ca claim frequecy
Chapter 7 Methods of Finding Estimators
Chapter 7 for BST 695: Special Topics i Statistical Theory. Kui Zhag, 011 Chapter 7 Methods of Fidig Estimators Sectio 7.1 Itroductio Defiitio 7.1.1 A poit estimator is ay fuctio W( X) W( X1, X,, X ) of
.04. This means $1000 is multiplied by 1.02 five times, once for each of the remaining sixmonth
Questio 1: What is a ordiary auity? Let s look at a ordiary auity that is certai ad simple. By this, we mea a auity over a fixed term whose paymet period matches the iterest coversio period. Additioally,
Baan Service Master Data Management
Baa Service Master Data Maagemet Module Procedure UP069A US Documetiformatio Documet Documet code : UP069A US Documet group : User Documetatio Documet title : Master Data Maagemet Applicatio/Package :
The Power of Both Choices: Practical Load Balancing for Distributed Stream Processing Engines
The Power of Both Choices: Practical Load Balacig for Distributed Stream Processig Egies Muhammad Ais Uddi Nasir #1, Giamarco De Fracisci Morales 2, David García-Soriao 3 Nicolas Kourtellis 4, Marco Serafii
Volatility of rates of return on the example of wheat futures. Sławomir Juszczyk. Rafał Balina
Overcomig the Crisis: Ecoomic ad Fiacial Developmets i Asia ad Europe Edited by Štefa Bojec, Josef C. Brada, ad Masaaki Kuboiwa http://www.hippocampus.si/isbn/978-961-6832-32-8/cotets.pdf Volatility of
The Power of Both Choices: Practical Load Balancing for Distributed Stream Processing Engines
The Power of Both Choices: Practical Load Balacig for Distributed Stream Processig Egies Muhammad Ais Uddi Nasir #1, Giamarco De Fracisci Morales 2, David García-Soriao 3 Nicolas Kourtellis 4, Marco Serafii
Lesson 15 ANOVA (analysis of variance)
Outlie Variability -betwee group variability -withi group variability -total variability -F-ratio Computatio -sums of squares (betwee/withi/total -degrees of freedom (betwee/withi/total -mea square (betwee/withi
Week 3 Conditional probabilities, Bayes formula, WEEK 3 page 1 Expected value of a random variable
Week 3 Coditioal probabilities, Bayes formula, WEEK 3 page 1 Expected value of a radom variable We recall our discussio of 5 card poker hads. Example 13 : a) What is the probability of evet A that a 5
Discrete Mathematics and Probability Theory Spring 2014 Anant Sahai Note 13
EECS 70 Discrete Mathematics ad Probability Theory Sprig 2014 Aat Sahai Note 13 Itroductio At this poit, we have see eough examples that it is worth just takig stock of our model of probability ad may
An Efficient Polynomial Approximation of the Normal Distribution Function & Its Inverse Function
A Efficiet Polyomial Approximatio of the Normal Distributio Fuctio & Its Iverse Fuctio Wisto A. Richards, 1 Robi Atoie, * 1 Asho Sahai, ad 3 M. Raghuadh Acharya 1 Departmet of Mathematics & Computer Sciece;
hp calculators HP 12C Statistics - average and standard deviation Average and standard deviation concepts HP12C average and standard deviation
HP 1C Statistics - average ad stadard deviatio Average ad stadard deviatio cocepts HP1C average ad stadard deviatio Practice calculatig averages ad stadard deviatios with oe or two variables HP 1C Statistics
Chapter 7: Confidence Interval and Sample Size
Chapter 7: Cofidece Iterval ad Sample Size Learig Objectives Upo successful completio of Chapter 7, you will be able to: Fid the cofidece iterval for the mea, proportio, ad variace. Determie the miimum
5.4 Amortization. Question 1: How do you find the present value of an annuity? Question 2: How is a loan amortized?
5.4 Amortizatio Questio 1: How do you fid the preset value of a auity? Questio 2: How is a loa amortized? Questio 3: How do you make a amortizatio table? Oe of the most commo fiacial istrumets a perso
Forecasting. Forecasting Application. Practical Forecasting. Chapter 7 OVERVIEW KEY CONCEPTS. Chapter 7. Chapter 7
Forecastig Chapter 7 Chapter 7 OVERVIEW Forecastig Applicatios Qualitative Aalysis Tred Aalysis ad Projectio Busiess Cycle Expoetial Smoothig Ecoometric Forecastig Judgig Forecast Reliability Choosig the
Definition. A variable X that takes on values X 1, X 2, X 3,...X k with respective frequencies f 1, f 2, f 3,...f k has mean
1 Social Studies 201 October 13, 2004 Note: The examples i these otes may be differet tha used i class. However, the examples are similar ad the methods used are idetical to what was preseted i class.
Agenda. Outsourcing and Globalization in Software Development. Outsourcing. Outsourcing here to stay. Outsourcing Alternatives
Outsourcig ad Globalizatio i Software Developmet Jacques Crocker UW CSE Alumi 2003 [email protected] Ageda Itroductio The Outsourcig Pheomeo Leadig Offshore Projects Maagig Customers Offshore Developmet
The Power of Free Branching in a General Model of Backtracking and Dynamic Programming Algorithms
The Power of Free Brachig i a Geeral Model of Backtrackig ad Dyamic Programmig Algorithms SASHKA DAVIS IDA/Ceter for Computig Scieces Bowie, MD [email protected] RUSSELL IMPAGLIAZZO Dept. of Computer
CS103A Handout 23 Winter 2002 February 22, 2002 Solving Recurrence Relations
CS3A Hadout 3 Witer 00 February, 00 Solvig Recurrece Relatios Itroductio A wide variety of recurrece problems occur i models. Some of these recurrece relatios ca be solved usig iteratio or some other ad
Page 1. Real Options for Engineering Systems. What are we up to? Today s agenda. J1: Real Options for Engineering Systems. Richard de Neufville
Real Optios for Egieerig Systems J: Real Optios for Egieerig Systems By (MIT) Stefa Scholtes (CU) Course website: http://msl.mit.edu/cmi/ardet_2002 Stefa Scholtes Judge Istitute of Maagemet, CU Slide What
