Designing Virtualization-Based Infrastructure for Providing On-demand HPC Software-based Service
Résumé
The emerging of Internet-based computing, namely cloud computing, has increased the possibility of sharing remote resource. On the other hand, while the usefulness and how to use grid computing for sharing hardware resource have been well studied, such studies do not seem to be available concerning software. We propose in this paper an on-demand service model to share software within grids. Targeting performance, flexibility and resource use efficiency, we discuss virtualization-enabled features to suggest an infrastructure model, and an implementation approach based on service execution within virtual machines (VM). The study show that, in spite of the potential performance overhead, on which conventional wisdom has based its unsuitability for HPC, virtualization can enable relevant features concerning on-demand cloud-like service. The possibility to build flexible and highly-reconfigurable infrastructure, and the possibility to easily achieve efficiency-driven resource scheduling are examples. Based on an OpenNebula infrastructure along with Xen as VM Monitor backend, we implement a prototype, and carry out experiments concerning performance using the parsec benchmark. Our results show that, with suitable tuning, VM can achieve near native performance, even when applications run onto single node-hosted concurrent VMs.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...