Browsing by Author "Vahvanen, Eetu"

Now showing 1 - 1 of 1

Multi-objective Computing Cluster Scheduling and Resource Reservation Optimization
(2020-05-08) Vahvanen, Eetu
Perustieteiden korkeakoulu | Bachelor's thesis
This research is a literature review that concludes multi-objective computing cluster optimization. Computing clusters have been developed since 1970s but the significance has risen in the global drive for digitalization. Modern clusters are generic, heterogenic systems that are capable of heavy computing. Their computing power is enormous compared to customer computers even though the components work in similar ways. This vast computing power is useful for at least research purposes, weather forecasting, and material innovations. The goal of this study is to analyze and conclude important techniques used for optimizing clusters. Study focuses on job scheduling and resource reservation which are essential factors in optimization. To further understand problems regarding optimization the study discusses three perspectives: configuration, system characteristics, and performance metrics. Clusters are compositions of several computers each capable for independent computing work. These computational units are called nodes which are interconnected in a very fast local network. The huge number of nodes modern computing clusters have call for a need for careful optimization. The main results of this study are as follows. Optimization regarding cluster resource reservation is necessary for the stability and utilization of the system. This enables fast execution of even heavier jobs. The visible part of optimization for a user is software that is responsible for schedules and dividing work to nodes. This part of a cluster is often called its real operating system for its central role in the system. Modern supercomputers, which are important use cases for clusters, are typically optimized with a FIFO algorithm. Other techniques, such as backfilling, are used to improve performance and resource utilization. The processing is typically done in batches where the users specify their needs. Other clusters, such as data centers, may use techniques capable of real-time optimization, and such techniques are actively researched.