On Advantages of Scheduling Using Genetic Fuzzy Systems

Abstract

Abstract. In this paper, we present a methodology for automatically generating online scheduling strategies for a complex scheduling objective with the help of real life workload data. The scheduling problem includes independent parallel jobs and multiple identical machines. The objective is defined by the machine provider and considers different priorities of user groups. In order to allow a wide range of objective functions, we use a rule based scheduling strategy. There, a rule system classifies all possible scheduling states and assigns an appropriate scheduling strategy based on the actual state. The rule bases are developed with the help of a Genetic Fuzzy System that uses workload data obtained from real system installations. We evaluate our new scheduling strategies again on real workload data in comparison to a probability based scheduling strategy and the EASY standard scheduling algorithm. To this end, we select an exemplary objective function that prioritizes some user groups over others.

Greedy scheduling with custom-made objectives

“…But if some of the restrictions regarding schedule feedback and timing are lifted then significantly better schedules can be produced that come close to the offline optimum, see, for instance, Franke et al (2006) who used Genetic Fuzzy methods to generate online scheduling methods. But these algorithms cannot be applied to the Grid problem described in Sect.…”

Cooperative negotiation and scheduling of scientific workflows in the collaborative climate community data and processing grid

“…That is, for every submitted workflow, a separate instance of the scheduling component can be created. This allows very flexible setups, including the assignment of different strategies to different user groups [15]. It is even possible to replace the applied scheduling strategy during runtime within an active workflow's execution process, such that system-state dependent decision strategies are feasible.…”

Job Mapping and Scheduling on Free-Space Optical Networks

“…The job running time is specified according to the Run Time. Figure 4 shows the CDF of the number of the requested CPUs in the workload log, which spreads from N = 1 to N = 2 8 . Around 82.1% of the jobs request N ≤ 2 4 CPUs.…”

“…In this case, we judge if doing so provides a benefit of shorter response time for the job when compared to scheduling the job later until enough resources are released for executing the job. Moreover, for preventing one job occupying some resources too long (e.g., four times larger than its original execution time) or beyond the requested runtime, we also use EASY Backfilling [8] to find an allocation for one of the following smaller jobs in the wait queue to the currently idle resources while ensuring that the first queuing job is not further delayed due to the backfilling.…”

“…Within the new algorithm, we consider an adaptive allocation for each user job such as packing expansion and packing shrinkage using the Downey model [7] as the job speedup model for the parallel jobs, which has been widely used to determine the relation of the job size and its corresponding execution time. We also use EASY Backfilling [8] to find an allocation for one of the following smaller jobs in the wait queue in case the first queuing job cannot be dispatched for its larger size than the idle resources. Meanwhile, we ensure that the first queuing job is not further delayed due to the backfilling.…”