A Joint Power Efficient Server and Network Consolidation approach for virtualized data centers

Abstract

Cloud computing and virtualization are enabling technologies for designing energy-aware resource management mechanisms in virtualized data centers. Indeed, one of the main challenges of big data centers is to decrease the power consumption, both to cut costs and to reduce the environmental impact. To this extent, Virtual Machine (VM) consolidation is often used to smartly reallocate the VMs with the objective of reducing the power consumption, by exploiting the VM live migration. The consolidation problem consists in finding the set of migrations that allow to keep turned on the minimum number of servers needed to host all the VMs. However, most of the proposed consolidation approaches do not consider the network related consumption, which represents about 10–20% of the total energy consumed by IT equipment in real data centers. This paper proposes a novel joint server and network consolidation model that takes into account the power efficiency of both the switches forwarding the traffic and the servers hosting the VMs. It powers down switch ports and routes traffic along the most energy efficient path towards the least energy consuming server under QoS constraints. Since the model is complex, a fast Simulated Annealing based Resource Consolidation algorithm (SARC) is proposed. Our numerical results demonstrate that our approach is able to save on average 50% of the network related power consumption compared to a network unaware consolidation.

Introduction

One of the main objectives in cloud computing is to find the right compromise between the cost-efficiency of the underlying infrastructure and the Quality of Service (QoS) as perceived by the users running their virtualized applications [1]. In the cloud computing paradigm there is an emerging green computing [2] awareness, which is strictly related to the cost-efficiency in the utilization of the physical resources: the aim is to reduce the power consumption and achieve an appropriate level of energy-efficiency. Big data centers, along with the adoption of the virtualization technology, are increasingly experiencing the need to reduce the energy consumption, because of both the environmental pollution and the economic concern. Hence, the interest of the research community is moving towards the introduction of metrics that allow to evaluate how energy efficient the resource utilization is. In the literature, techniques that try to increase the resource utilization of data centers have been massively studied from the energy view-point. One of them is the VM consolidation, whose objective is to minimize the number of physical servers necessary to host a set of VMs. By leveraging the VM live migration [3], the allocation of VMs to physical nodes can be dynamically adjusted to achieve different goals such as load balancing, avoiding hot spots or hibernating under-utilized servers.

Many VM consolidation techniques do not take into account networking issues within a data center. Greenberg et al. [4] shows, that the network consumes around 20% of the total energy in a data center. Large data centers are providing cloud enabled services and applications, which typically consist of multiple cooperating VMs that need to exchange large data volumes over the data center network. Classical network infrastructures, based on the spanning tree topology, have been proved to suffer from severe performance limits, such as network throughput, equipment capital expenditure (CapEx), network diameter and so on. In [5] a new parameter, namely the Network Power Effectiveness (NPE), is introduced to evaluate the network efficiency as the ratio between the aggregate throughput and the total power consumption. Along with novel architectures, techniques for increasing energy efficiency in the network are also emerging. Commercial switches allow putting network interfaces in sleep-mode when no traffic is flowing, and waking-up a link only when a packet arrives. This is useful since several papers argue that the traffic load is very day-time dependent in data centers. For example, [6] shows that the average link utilization in the connections to the aggregate switches is only 8% of the capacity for 95% of the time, while in the core layer the utilization is between 20% and 40%. Therefore, it is possible to dynamically switch-off under-utilized network devices and links in some time intervals in order to save energy. Network protocols can also benefit from energy efficiency. For example, the IEEE 802.3az [7] standard is aimed at achieving energy reduction in the Ethernet based communications by activating a link only when real data is being sent.

In [8], we proposed a new MILP model and a heuristic to consolidate a virtualized data center by reallocating VMs on the smallest subset servers in order to minimize the total energy consumption due to compute resources. The key contribution of this work is a new mathematical model extending [8], that also considers the network characteristics in order to jointly optimize the VM placement and energy efficient network routing. By knowing the current VM placement and the network routing for sending VM-to-VM traffic, we consider the energy efficiency profiles of servers and networking devices in order to find the VM migrations and network paths that jointly minimize the server and network power consumption, powering down unused switches and link ports which are not necessary for routing the traffic demands. As the proposed model is complex and very hard to solve, we develop a fast Simulated Annealing based Resource Consolidation (SARC) heuristic. The parameter, called migration desirability, is used in the perturbation phase of the heuristic: it allows us to balance between the resource utilization of the physical servers, the power efficiency and the impact the migrations may have on the power-aware established network paths. We show that our heuristic is able to save on average 50% of the total network power consumption when compared to the heuristic that simply consolidates the active servers.

The paper is structured as follows: in Section 2, the related work is discussed. In Section 3, the problem is formulated and the objective is explained, while in Section 4, a Mixed Integer Programming Model is described. In Section 5, the heuristic is presented and Section 6 shows the experimental evaluation. Finally, the paper concludes in chapter VII.