Mitigating Performance Unpredictability in Heterogeneous Clouds

Boris Teabe (1, 2) , Alain Tchana (1) , Daniel Hagimont (1, 2)

1 IRIT-SEPIA - Système d’exploitation, systèmes répartis, de l’intergiciel à l’architecture
2 Toulouse INP - Institut National Polytechnique (Toulouse)

The speed of a device may vary since (i) IaaS hardware infrastructures are increasingly heterogeneous and (ii) devices often have a dynamically adjusted speed in order to adapt their energy consumption according to the load. This paper addresses SLA enforcement in a IaaS which includes devices whose speed vary. We show that resource management should rely on an absolute value SLA specification (i.e., a performance metric which is independent from the device speed) and a dynamic translation of this SLA into actual allocations according to the device speed. Surprisingly, while disk or network resource allocations already integrate such a scheme, CPU does not. We propose a CPU resource management system which implements absolute CPU allocation and dynamically translates it into actual CPU allocations according to CPU speed. We demonstrate and evaluate the benefits of this resource management system.

Mots clés: Cloud; Heterogeneous architecture; DVFS

Domaines: Réseaux et télécommunications [cs.NI]; Système d'exploitation [cs.OS]; Architectures Matérielles [cs.AR]; Systèmes embarqués

Boris Teabe, Alain Tchana, Daniel Hagimont. Mitigating Performance Unpredictability in Heterogeneous Clouds. 13th IEEE International Conference on Services Computing (SCC 2016), Jun 2016, San Francisco, CA, United States. pp. 593-600. ⟨hal-01782589⟩