Description of the adaptive resource management problem, cost functions and performance objectivesDocUID: 1995-004
Author: G. Georgiannakis, Catherine Houstis, Sarantos Kapidakis, Maria Karavassili, Christos Nikolaou, Alexandros Labrinidis, Manolis Marazakis, Evangelos Markatos, M. Mavronicolas, S. Chabridon, E. Gelenbe, E. Born
Abstract: In LYDIA, we are interested in studying the problem of adaptive resource management specifically for distributed (message passing) and virtual shared memory architectures and workstations clusters. We describe these systems in more detail in chapter 2. We also concentrate our attention on specific applications: transaction management (not sequential or parallel query processing), numerically intensive parallel computations, multimedia and intelligent agent systems. We do not address the problem of adaptive memory or bandwidth management. These applications generate a workload and its monitoring, tracing, measurement and characterization is crucial when selecting the appropriate resource management policy. These issues are addressed in chapter 3. In this deliverable we survey recent work reported in the literature on adaptive management policies of the computational resources, when used in architectures and for applications that are of interest to LYDIA. These policies lie in a continuum, one end of which is task assignment and the other end is dynamic load sharing in the following sense: In task assignment, we have almost all the information we need for the a priori static assignment of tasks. The algorithms for task assignment are complex, off-line and address assignment to processors of rather large chunks of work (relatively long time scales). At execution time, the system waists almost no overhead to decide which unit of work executes next on which processor. It just follows a preestablished script. At the other end of the spectrum is fully dynamic load sharing where nothing is known a priori but decisions are made on the fly by the system based on some appropriate load indices. Scheduling lies between assignment and fully dynamic load sharing. Scheduling means that some information is known a priori, according to which, some possibly complex scheduling rules have been designed for the system to use. At run time these rules are combined with other, on the fly information, to make a scheduling decision. In addition, policies for the management of computational resources are greatly influenced by the structure of the units of work to be managed. We address the problems arising from managing sequential units of work in chapter 5 and those appearing from the management of parallel units of work in chapter 6. Finally, we present our research program in LYDIA in chapter 7.
Published In: Institute of Computer Science, FORTH, Technical Report No. 130
Year Published: 1995
Project: Others Subject Area: Others
Publication Type: Technical Report