High performance computing in the cloud
Year: 2017
Type: dissertation
Abstract: In recent years, the interest in both scientific and business workflows has increased. A workflow is composed of a series of tools, which should be executed in a predefined order to perform an analysis. Traditionally, these workflows were executed in a manual way, sending the output of one tool to the next one in the analysis process. Many applications to execute workflows automatically, appeared recently. These applications ease the work of the users while executing their analyses. In addition, from the computational point of view, some workflows require a significant amount of resources. Consequently, workflow execution moved from single workstations to distributed environments such as Grids or Clouds. Data management and tasks scheduling are required to execute workflows in an efficient way in such environments. In this thesis, we propose a cloud-based HPC environment, focusing on tasks scheduling, resources auto-scaling, data management and simplifying the access to the resources with software clients. First, the cloud computing infrastructure is devised, which includes the base software (i.e. OpenStack) plus several additional modules aimed at improving authentication (i.e. LDAP) and data management (i.e. GridFTP, Globus Online and CloudFuse). Second, built on top of the mentioned infrastructure, the TORQUE distributed resources manager and the Maui scheduler have been configured to schedule and distribute tasks to the cloud-based workers. To reduce the number of idle nodes and the incurred cost of the active cloud resources, we also propose a configurable auto-scaling technique, which is able to scale the execution cluster depending on the workload. Additionally, in order to simplify tasks submission to the TORQUE execution cluster, we have interconnected the Galaxy workflows management system with it, therefore users benefit from a simple way to execute their tasks. Finally, we conducted an experimental evaluation, composed by a number of different studies with synthetic and real-world applications, to show the behaviour of the auto-scaled execution cluster managed by TORQUE and Maui. All experiments have been performed by using an OpenStack cloud computing environment and the benchmarked applications correspond to the benchmarking suite, which is specially designed for workflows scheduling in the cloud computing environment. Cybershake, Ligo and Montage have been the selected synthetic applications from the benchmarking suite. GECKO and a GWAS pipeline represent the real-world test use cases, both having a diverse and heterogeneous set of tasks.
Author: Óscar Torreño Tirado
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2883661837
