Scheduling Divisible Loads to Optimize the Computation Time and Cost

Abstract

Efficient load distribution plays an important role in grid and cloud applications. In a typical problem, a divisible load should be split into parts and allocated to several processors, with one processor responsible for the data transfer. Since processors have different speed and cost characteristics, selecting the processors order for the transmission and defining the chunk sizes affect the computation time and cost. We perform a systematic analysis of the model analysing the properties of Pareto optimal solutions. We demonstrate that the earlier research has a number of limitations. In particular, it is generally assumed that the load should be distributed so that all processors have equal completion times, while in fact there often exists a dominating schedule with nonsimultaneous finishing times of the processors. Moreover, fixing the processor sequence in the non-decreasing order of the cost-characteristic may be appropriate only for Paretooptimal solutions with relatively large deadlines; optimal schedules for tight deadlines may have a different order of processors. We conclude with an efficient algorithm for finding the time-cost tradeoff.

Scheduling divisible loads with time and cost constraints

“…the problem is solvable in O(m 3 ) time even in the bicriteria setting (Shakhlevich 2013). As we show in this paper, the general case with arbitrary…”

Adapting Market-Oriented Policies for Scheduling Divisible Loads on Clouds

“…TheOPRisasuperiorpartitioningstrategybasedondivisibleloadtheory(DLT)whichhas beenintroducedinBharadwaj,Ghose,andRobertazzi,(2003)aswellasLin,Deogun,andGoddard (2007.Itstatesthattheoptimalexecutiontimeisobtainedwhenallnodesallocatedtoataskcomplete theircomputationatthesametime.Forcomparison,abaselineapproachtopartitionaworkloadN equal-sizedsubtaskisadopted.ThisapproachisnamedastheEqual Partitioning Rule (EPR).Each partitioningmethodhasdifferentadvantages;computationtimeandcostwouldbeexpectedtovary withadifferentstrategy.ThelastdecisionisWorkerSelectionStrategytodeterminetheworkersequenceorder.Selectionof workersincloudcomputingcurrentlyisdoneusingasimpleround-robinscheme,andtheimportance ofworkerselectionhasbeenhighlightedintheliterature(Pokahr&Braubach,2016;Shakhlevich, 2013;Yuan,Bi,Tan,&Li,2016).Toobservetheimpactofthisadditionalstrategy,threesequence ordersofworkerswereexamined(Majid&Chuprat,2018).Ifnoschedulecanbefoundtosatisfy thejob'stimingrequirementalthoughenoughvirtualmachineshavebeenadded,thejobwillbe rejected.TheresultofthesimulationisthoroughlyanalyzedintheSectiononExperimentalResult.PROPOSeD MORTDLT MODeLOur proposed Market-Oriented Real-time Divisible Load Theory scheduling algorithm named MORTDLTallowsausertochoosemorepowerfulandfastresourcesorlesspowerfulandslower resourcesaccordingtotheirpreferenceintermsoftheircostbudgetandjobpriority(speedofresourcesislessthan50MIPS) S F =fastspeed(speedofresourcesismorethan50MIPS)…”