Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.589845
Title: Addressing parallel application memory consumption
Author: Perks, Oliver F. J.
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 2013
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Recent trends in computer architecture are furthering the gap between CPU capabilities and those of the memory system. The rise of multi-core processors is having a dramatic effect on memory interactions, not just with respect to performance but crucially to capacity. The slow growth of DRAM capacity, coupled with configuration limitations, is driving up the cost of memory systems as a proportion of total HPC platform cost. As a result, scientific institutions are increasingly interested in application memory consumption, and in justifying the cost associated with maintaining high memory-per-core ratios. By studying the scaling behaviour of applications, both in terms of runtime and memory consumption, we are able to demonstrate a decrease in workload efficiency in low memory environments, resulting from poor memory scalability. Current tools are lacking in performance and analytical capabilities motivating the development of a new suite of tools for capturing and analysing memory consumption in large scale parallel applications. By observing and analysing memory allocations we are able to record not only how much but more crucially where and when an application uses its memory. We use use this analysis to look at some of the key principles in application scaling such as processor decomposition, parallelisation models and runtime libraries, and their associated effects on memory consumption. We demonstrate how the data storage model of OpenMPI implementations inherently prevents scaling due to memory requirements, and investigate the benefits of different solutions. Finally, we show how by analysing information gathered during application execution we can automatically generate models to predict application memory consumption, at different scale and runtime configurations. In addition we predict, using these models, how implementation changes could affect the memory consumption of an industry strength benchmark.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.589845  DOI: Not available
Keywords: QA76 Electronic computers. Computer science. Computer software
Share: