Performance-targeted Resource-aware Task Scheduling for Heterogeneous Platforms

Examensarbete för masterexamen

Please use this identifier to cite or link to this item: https://hdl.handle.net/20.500.12380/255113
Download file(s):
File Description SizeFormat 
255113.pdfFulltext2.56 MBAdobe PDFView/Open
Type: Examensarbete för masterexamen
Master Thesis
Title: Performance-targeted Resource-aware Task Scheduling for Heterogeneous Platforms
Authors: Fahlgren, Henrik
Rohlin, Agnes
Abstract: Embedded environments often come with strict energy requirements, especially in mobile devices. Heterogeneous architectures are often used in these systems as they provide resources useful in different scenarios. However, introducing more resources requires proper scheduling to be able to utilize the resources efficiently. XiTAO is a resource-aware runtime which dynamically allocates appropriate resources to provide interference-free execution. This thesis aims to extend the XiTAO runtime to consider heterogeneous scheduling as it currently only considers scheduling for a homogeneous platform. For this thesis, only architectures with heterogeneous cores using the same instruction set architecture are considered. Specifically, a Huawei processor with ARM big-LITTLE architecture nodes mounted on a Hikey960 board is used. After surveying related work, four scheduling extensions are presented and evaluated in the thesis. Two extensions target the critical path of the application of which one schedules tasks on the critical path on predefined big cores while the other uses a history-based method of finding the most suitable cores. The third extension uses a history-based method to recognize which tasks will get the greatest performance increase from running on a big core compared to a LITTLE core. The fourth extension changes the amount of resources given to each task dynamically by observing the load and history of the system together with a moldability feature of XiTAO. The last extension can be applied simultaneously as the other scheduling extensions or individually. The scheduling extensions were evaluated with synthetic benchmarks with three different kernels, highlighting specific scenarios. The results show a speedup up to 29% for a randomized directed acyclic graph with our extensions compared the original runtime scheduling.
Keywords: Informations- och kommunikationsteknik;Data- och informationsvetenskap;Information & Communication Technology;Computer and Information Science
Issue Date: 2018
Publisher: Chalmers tekniska högskola / Institutionen för data- och informationsteknik (Chalmers)
Chalmers University of Technology / Department of Computer Science and Engineering (Chalmers)
URI: https://hdl.handle.net/20.500.12380/255113
Collection:Examensarbeten för masterexamen // Master Theses



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.