Aristotle University Of Thessaloniki (AUTH)
Οptimization of scheduling in workflow management systems using reinforcement learning
Abstract
dc:descriptionA workflow is defined as the execution of a sequence of tasks, the order of which is determined by data interdependencies and the target outcome. The execution of workflows can be managed by a Workflow Management System (WMS) which undertakes the scheduling of tasks, handling of failures and monitoring health status. Scheduling workflows involves mapping tasks to available execution sites in respect to a cost function that optimizes an objective such as the total execution time. In the era of multi-core architectures, accelerated computing is largely driven by scaling across distributed resources. For a WMS, identifying and exploiting opportunities for parallelism becomes a critical necessity. At the same time, distributed execution introduces new challenges as workflow execution should be agnostic to the software and hardware on the execution sites. Additionally, machine learning based approaches to scheduling optimization have not been explored to the same extent as heuristic algorithms. The aim of this dissertation is to exploit opportunities for parallelism and leverage machine learning in scheduling scientific workflows across distributed heterogeneous execution sites. The outcome of this endeavor is Hermes, a novel WMS capable of exploiting opportunities for parallelism, supporting distributed execution and leveraging machine learning to profile tasks and computational resources and perform scheduling decisions. Although capable of supporting any workflow, Hermes is evaluated with bioinformatics workflows as they comprise most of its user base. A methodology is proposed for addressing data and pipeline parallelism by compartmentalizing computational flows into pipelines with minimal data interdependencies. The methodology is applicable to workflows with multiple sequential processes that can be applied individually on fragments of the original input. A comparison against a high-performance framework for distributed computing is performed evidencing reduced execution time and costs. Hermes provides an environment agnostic solution to seamlessly create, share and reproducibly execute scientific workflows. Distributed execution is made possible via container-based virtualization technologies that accomplish standardization of the execution environment. Hermes employs binary classification models to predict task failures and regression models for predicting task runtime given variable task input sizes and execution sites of varying computational capabilities. The results confirm that useful and directional metrics can be achieved that can influence scheduling decisions. The use of seq2seq neural architectures is proposed for modelling task to execution site scheduling decisions. Expected failure probabilities and runtimes of combinations of tasks mapped to execution sites are the inputs to the neural architecture and the outputs are the selected mappings. As generating high quality labeled data can be prohibitively expensive when dealing with NP-Hard problems, supervised learning can be challenging. For this reason, a reinforcement learning training approach is adopted. Hermes implements an environment for distributed workflow execution simulation which can simulate not only task execution times but also task failures and the cost of network communications. The trained neural network is evaluated at performing scheduling decisions when executing real workflows. As demonstrated, it significantly outperforms traditional scheduling algorithms used in WMSes in execution time, failed tasks, and size of network traffic.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kintsakis, Athanassios
- Κιντσάκης, Αθανάσιος
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/45467