Themis: Economy-based Automatic Resource Scaling for Cloud Systems

High performance computing (HPC) infrastructures are used to execute increasingly complex scientific applications with time-varying resource requirements. A key challenge for infrastructure providers is to distribute resources to such applications so that application performance objectives are met w...

Full description

Saved in:
Bibliographic Details
Main Authors: Costache, S., Parlavantzas, N., Morin, C., Kortas, S.
Format: Conference Proceeding
Language:English
Subjects:
cloud computing
Computational modeling
Dynamic scheduling
Feedback control
Measurement
Pricing
Resource management
spot markets
Virtual machining
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:High performance computing (HPC) infrastructures are used to execute increasingly complex scientific applications with time-varying resource requirements. A key challenge for infrastructure providers is to distribute resources to such applications so that application performance objectives are met while guaranteeing a high infrastructure utilization. Most of existing solutions provide limited support for meeting application objectives and can lead to inefficient resource usage. In this paper we propose Themis, a system that uses an economic-based approach to automatically allocate resources to the applications that need them the most. Themis relies on a proportional-share auction that ensures fair differentiation between applications while maximizing the resource utilization. To support meeting application performance objectives, Themis provides generic scaling policies based on feedback control loops. These policies adapt the resource demand to the current infrastructure conditions and can be extended for different application types. We have evaluated the performance of the system through simulations, and the results show that Themis can effectively meet application performance objectives while optimizing infrastructure's utilization.
DOI:10.1109/HPCC.2012.56