PARVIS - Performance mAnagement of VIrtualized Systems
Virtualization of physical resources is receiving great interest both from industry and academia as a way to develop smart and flexible cloud infrastructures. Virtual machine monitors (VMMs) provide performance differentiation and performance isolation to competing running virtual machines (VMs). Each VM has access to a time-varying fraction of the physical server capacity defined by the VMM resource allocation parameters (i.e., VMWare shares or Xen weights), which can be updated in few milliseconds without introducing any system overhead. Hence, this feature can be adopted effectively to control and adapt the application performance to incoming workload on very fine-grained time scales.
However, virtualization introduces new problems related to resource multiplexing and scheduling of multiple VMs: the performance modelling of virtualized environments is challenging, as the impact of the choice of the VMM scheduler, its parameters and I/O management overhead is still only partially understood. Furthermore, the autonomic management implemented today by virtualization products (e.g., VMWare DRS or Virtuoso VSched) aims at equalizing the use of resources but cannot provide any Quality of Service (QoS) guarantee.
The problem is made particularly challenging by the high variability of the incoming workload, which can change, by several orders of magnitude within a single business day. Techniques able to model and control the system at very ﬁne-grained time scales and in transient conditions have to be developed.
The aim of this research is to develop novel resource allocation policies for smart dependable virtualized cloud infrastructures via an interdisciplinary approach.
The research proposes the adoption of performance evaluation and optimization methods for the long-term (e.g., half an hour) management of the physical infrastructure, while system identification and control engineering methods are used to derive load-dependent black-box models of virtualized cloud systems from experimental data and to design short-term control systems (e.g., operating at minute time scale).
This model-based approach allows an accurate and effective management of the running applications providing performance and dependability guarantees.
- D. Ardagna, B. Panicucci, M. Trubian, L. Zhang Energy-Aware Autonomic Resource Allocation in Multi-tier Virtualized Environments. IEEE Transactions on Services Computing. To Appear.
- D. Ardagna, M. Tanelli, M. Lovera, L. Zhang. Black-box Performance Models for Virtualized Web Service Applications . Proceedings of the 1st Joint WOSP/SIPEW International Conference on Performance Engineering. WOSP/SIPEW2010. ACM DL, 153-163. San Jose, CA, USA.
- B. Addis, D. Ardagna, B. Panicucci , L. Zhang. Autonomic Management of Cloud Service Centers with Availability Guarantees. In Cloud 2010 Proceedings. To appear.
- GAME-IT Green Active Management of Energy in IT systems