Dynamic and Selfadaptive ResourceManagement: 2ETHER Operating Environment
Fabrice Muller · 2007
Thedevelopment ofreconfigurable devices thatcould makethemselves domain-specialized atrun timeisbecoming more andmore common. Future reconfigurable architecture will havethese computing devices asbasic blocks, andreconfigurable architecture couldmakeassemblies ofthesedevices, on thefly, to executeconcurrentapplications. The migration from completely generic lookuptables andhighly connected routing fabrics toselfadaptive specialized coarse-grain reconfigurable devices andvery structured communication resourcespresents designers withtheproblem ofhowtobestcustomize thesystem basedupon anticipated usage.Thenthereisa needofnotonly exploiting parallelism fromapplications atmicro-thread level, dynamically, butsystemalsostarves fora dynamicandself adaptive middleware to schedule thesemicro-threads on thousands ofsuchcomputing devices. Thispaper focuses atthe problemofdynamicallocation andscheduling ofresourcesto numbersofapplications on sucharchitecture. Keywords-Resource Allocation; dynamicscheduling; self adaptive taskstructuring. I. INTRODUCTION Inorder toprovide highperformance computation power toservetheincreasing needoflarge applications, people strive to improve a single machine's capacity or construct a distributed systemcomposed ofa scalable setofmachines. Compared totheformer, wheretheimprovement ismainly up tothehardware technology development, theconstruction of distributed systemsforresource collaboration ismore complex. Someofwell-known existing distributed systems composedof heterogeneous resources are Condor(4), NetSolve(5), Nimrod(6), Globus andtheGrid(7) computation environment. Sabinet al (9)propose a centralized metasheduler which usesbackfill toschedule parallel jobsin multiple heterogeneous sites. Similarly, Aroraetal(10) present a completely decentralized, dynamicandsender- initiated scheduling andloadbalancing algorithm fortheGrid environment. Allthese approaches don't dealwithapplication modelwhereconcurrent threads arecreated andmanaged at run.Thesemethods donottarget future architecture where eachresourceofa processorhascapability ofself optimizing andinterconnects withother resourcestoformassemblies to execute concurrent application. IETHERI systemishierarchical bothatfunction and architecture level. Atfunction level, application iswritten whichselfadapts itself towellsuitwiththeapplication objective andtocope withdynamic changes happening in environment. Applications arequite dynamic innature where concurrent threads are instanced dynamically according to numberofresourcesofsystem. Architecture ofthesystemis nottraditional asitisnotsingle unitofcomputation. Itisa network ofgiven numberofSANEs.A SANEisself adaptive networked entity which canself optimize itself. Inclassical scheduling theory ithasbeencommonly assumed that ataskrequires only one processoratatimefor itsprocessing. Howeverformany practical problems this assumption isnotvalid. A taskrequires task-dependent (worst case execution time)numberof processors(minimum) simultaneously forits processing.