Inactivity Benchmarking

Wictor Lund, Johan Lilius

Research output: Chapter in Book/Conference proceedingPublished conference proceedingScientificpeer-review

Abstract

Computational workloads that run out of work need some way to signal the intent to be inactive. With a proper inactivity signaling, the battery life can be extended, cloud costs can be reduced and the performance can be improved. In this paper, we study some properties of the inactivity signaling capabilities on virtualized platforms. The main contribution of the paper is a presentation of the inner workings of Spurg-Bench - a benchmarking tool developed by the author. The Spurg-Bench generates a load consisting of computation and inactivity. A secondary contribution is a small study on the inactivity behaviour of two different operating system (OS) kernels.Spurg-Bench generates a specified load on a system according to the utilization of the operations per second (OPS). By using the utilization notion of the load, Spurg-Bench generates a load occupying a fraction of the available processor time without any throughput guarantees. With the OPS notion, the generated load has a given throughput but has no guarantees on how much processor time it uses.Spurg-Bench generates load in cycles by executing an operation repeatedly and becoming inactive for a while. The parameters of a cycle in Spurg-Bench are the number of the operations and the duration of the inactivity. To achieve the desired load, the parameters of Spurg-Bench are dynamically adjusted using a control algorithm which minimizes the difference between the current load and the specified load. Spurg-Bench differs from similar tools by running the workload in the same loop as the control algorithm. The in-loop approach is crucial to the inactivity benchmarking because it enables a full control over the duration and the starting time of the inactivity periods.

Original languageUndefined/Unknown
Title of host publication2018 International Conference on High Performance Computing & Simulation (HPCS)
EditorsWaleed W. Smari Khalid Zine-Dine
PublisherIEEE
Pages1028–1033
ISBN (Electronic)978-1-5386-7879-4
ISBN (Print)978-1-5386-7878-7
DOIs
Publication statusPublished - 2018
MoE publication typeA4 Article in a conference publication
EventInternational Conference on High Performance Computing & Simulation (HPCS) - 2018 International Conference on High Performance Computing & Simulation (HPCS)
Duration: 16 Jul 201820 Jul 2018

Conference

ConferenceInternational Conference on High Performance Computing & Simulation (HPCS)
Period16/07/1820/07/18

Keywords

  • Cloud computing
  • cloud computing

Cite this