The Research Of Trace-based Workload Modeling And Scheduling Optimization

With the rapid growth of High Performance and Cloud Computing, their workload scalegrows rapidly too. Although the much larger computing system can deal with the much largerworkload, it is not the efficient one. The efficient method is to make higher use of resourcethrough efficiency scheduling. In High Performance and Cloud Computing, scheduling is therelational mapping between workload and resource. Therefore, the trace analysis andmodeling can identify the performance deficiency of scheduling, thus we can improvescheduling and evaluate the performance of the improved scheduling. To this end, ourcontributions include:(1)Firstly we propose and construct a universal framework for workload modeling fromrelated work. The universal framework input and standardize raw trace, classify the workloadinto rigid job and moldable job, then analyze the workload characteristics of raw traceaccording to requirements and find suitable distributions for these characteristics, thencalculate the parameters of the distributions for the trace, finnaly construct a workload modelwhich can be used to generate uniform distributed workload with the trace. Secondly wemodel VM CPU usage rate trace and Biological Gene Sequencing (BGS) trace through theuniversal framework, and assessment results show that the two models can be used togenerate uniform distributed workload with the traces, which demonstrates the universality ofthe framwork.(2) For the non-preemptive of the rigid job, we propose and use one-dimensionalbin-packing to describe the relationship between the VM and the resource, and limit it withenergy consumption. Then we have proposed two proposals to improve the existing ones.1)Firstly, for the disadvantages of current VM Overload Decision Algorithms (ODA) and VMSelection Algorithms (VMSA), we propose a novel ODA based on VM CPU usage ratemodel, and an improved VMSA. The novel ODA use the mean and standard deviation of VMCPU usage rate to assess the host overload or not. if the host is overload, then the improvedVMSA select a suitable VM on the host to migrate by the minimum positive correlationcoefficient. The experiments show that the novel ODA and improved VMSA outperform thecurrent ODAs and VMSA. The current ones get3.84for Energy-Performance Tradeoff (EPT),but ours get1.28for EPT.2) Secondly, for the disadvantages of current heuristic frameworkfor VM consolidation, we propose a redesigned framework for VM consolidation. In theredesigned one, we propose a SLA Violation Decision Algorithm (SLAVDA) based on VMCPU usage rate model, and an improved Minimum Power Minimum Utilization policy (MPMU). SLAVDA can be used to assess a host SLA violation or not. If the host is SLAviolation, then we can migrate some selected VMs (selected by VMSA) on some suitable hostaccording to MPMU. We apply the VM CPU usage rate trace and VM CPU usage rate modelto evaluate the redesigned framework. Experimental results show that the redesignedframework outperforms the current one on energy consumption, SLA violation and EPT.(3) For the preemptive of the moldable job, we propose and use queuing theory todescribe the relationship between the job and the resource. Firstly we propose an AverageWeighted to calculate the User-Trust based on BGS model. Backfilling can predict VMruntime through User-Trust and VM’s require runtime (Trust). Then we evaluate theperformance of the Trust and existing proposal (Tsafrir) through the trace-driven simulation.Experimental results show that Trust outperforms Tsafrir on accuracy, average wait-time andbounded slowdown. Finally, the performance evaluation also shows that Trust outperformsTsafrir on average wait-time and bounded slowdown with BGI trace and BGIModel.We have improved scheduling through the workload characteristics of workloadmodeling, and evaluated the performance of the improved scheduling. The experimentalresults show that the improved one outperforms the previous design. Meanwhile, we haveevaluated the improved scheduling with workload models and get uniform results with traces,which verify the feasibility of the models and the robustness of the improved scheduling. Infuture, we can port the improved scheduling to the reality and use it to analyze the onlinetrace and verify its reliability and robustness.