| Abstract [eng] |
The providers of Hybrid Cloud based platforms must solve the challenge of workload scheduling, meanwhile satisfying Service Level Agreements. Overview of scientific publications shows that costs for running idle resources increases significantly if scheduling is done irresponsibly. Motivated by this problem, task scheduling theory were adapted to maximize private Cloud utilization. Main property which lets any schedule customizations is task’s priority or maximum postpone time. In this work, scheduling also involves task’s run time, required virtual machine number and from hardware side – private Cloud capacity. In conjunction with scheduling comes restrictions like preventing high priority tasks from being sent to public Cloud. This leads to investigation of the workload data characteristics while restriction management refers to prediction model that generates future values of resource occupation or incoming workflow. To establish optimal schedule according to mentioned restrictions, objective function is needed which minimizes processor’s time rented from public Cloud. Research showed that fitting any standard mathematical distribution like Pareto, Weibull or Gama to characteristics of the real-life workload data is a complicated task. Real-life data is frequently stochastic or modal and does not fit theoretical models, therefore empirical distributions are the best choice to use. In terms of workflow, the autocorrelation analysis confirmed a weekly seasonal pattern with decreasing traffic at weekends and similar tendency in workdays. However, despite seasonal component, more accurate workflow prediction using neural networks, does not prove to be very realistic. Lastly, hybrid Cloud model was created, in which particle swarm optimization were used to schedule tasks while taking into account basic workload characteristics, scheduling restrictions and resource occupation prediction. After comparing this type of scheduling with FiFo and “short tasks first” strategies it did show some increase in private Cloud utilization. However the best results there achieved after additionally adding tasks pre-execution. This means that no matter how good the schedule is, not all tasks fits into private Cloud. Consequently, these excessive tasks could be sent to public Cloud right away. Overall, task scheduling strategies implemented in this work allowed to increase private Cloud utilization and most of all – shortened an average tasks waiting time. |
