Editorial: Resource management of parallel and distributed systems with static scheduling: Challenges, solutions and new problems

Abstract

' While the term 'scheduling' has a certain meaning-that is. sequencing a set of jobs to a set of resources-it has been used interchangeably with the general resource management problem; this point is further elaborated later. ISHFAQ A H M A D CLASSIFICATION OF RESOURCE MANAGEMENT PROBLEMSSince there are many variants of the resource management problem (depending upon the characteristics of the application and the system as well as the assumption and objectives of the solution), a number of methods for classifying these problems have also been proposed [4]. One possible classification (Figure 1) is local versus global: local scheduling means allocating multiple processes to the time slices of a single processor; global scheduling means assignment of tasks to multiple processors with a system-wide performance objective. Since we are more interested in global scheduling, we can further classify the global scheduling depending upon whether the tasks are independent or related. Expanding the classification of multiple task scheduling, we can further classify the problem into static and dynamic. Static scheduling means that all of the allocation decisions are independent of the state of the system and are done before running the problem. In other words there is no change in the allocation decision after the problem has started its execution. Static scheduling is also called deterministic scheduling. If the scheduling decisions are made 'on the fly', depending upon the state of the system, scheduling is said to be dynamic. The disadvantage of static scheduling is that it does not change its behavior according to the state of the system. The disadvantage of dynamic scheduling is that it cannot tolerate excessive scheduling overhead because the schedule must be determined on the fly. The static scheduling itself exists in many forms. As a result, there are various terms that imply different meanings of resource management. Program partitioning refers to decomposing the problem into many tasks; this may be done irrespective of the target processor system, or it may be done by considering the specific system. The decomposition of the problem can also be of two types: domain decomposition which refers to the partitioning of input data, and controf decomposition which refers to the partitioning of the code. Clustering is the process of grouping tasks into clusters, while mapping is the process of assigning tasks or clusters of tasks to processors. Load balancing attempts to EDITORIAL 34 1 assign an equal amount of computational load to every processor in the system. Load sharing tries to optimize the performance of the system by transferring some of the load from heavily loaded processors to lightly loaded or idle processors. Sometimes, the terms load balancing and load sharing are used without any distinction. The term scheduling in a more restricted sense means allocation of the tasks of an application to the processors such that the running time of the entire application is minimized without violating the temporal depe...

Competitive Coevolutionary Multi-Agent Systems: The Application to Mapping and Scheduling Problems

“…Some of them use a heuristic search based on simulated annealing [7], genetic algorithms [14], or neural networks [18]. Some insight into currently used methods of mapping and scheduling can be found in the recent literature [1,2]. Most known heuristics are serial algorithms which are executed on sequential machines.…”

A software-based MPEG-4 video encoder using parallel processing

“…In a data parallel program, the issue of load balancing should be carefully addressed. Load balancing means equalization of the processors' workloads to minimize their idle times [12]. In MPEG-4 video, the size and location of each object may vary with time, and such behavior cannot be predicted beforehand.…”

Theory and practice in parallel job scheduling

“…See, for example the proceedings of three workshops [40], and a survey of a large number of proposed and implemented approaches [19]. It has become a distinct research topic from the largely unrelated, but better known problem of DAG scheduling [1]. DAG scheduling assumes that all tasks have fixed and specified dependencies, whereas job scheduling assumes that the jobs are mostly independent.…”