Efficient Server Provisioning and Offloading Policies for Internet Data Centers with Dynamic Load-Demand

Abstract

In data centers, traffic demand varies in both large and small time scales. A data center with dynamic traffic often needs to over-provision active servers to meet the peak demand, which incurs significant energy cost. In this paper, our goal is to reduce energy cost of a set of distributed Internet data centers (IDCs) while maintaining the quality of service of the dynamic traffic. In particular, we consider the outage probability as the QoS metric, where outage is defined as service demand exceeding the capacity. We require the outage probability at each IDC to be smaller than a predefined threshold. Our goal is thus to minimize total energy cost over all IDCs, subject to the outage probability constraint. We achieve the goal by dynamically adjusting server capacity and performing load shifting in different time scales. We propose three different load-shifting and joint capacity allocation schemes with different complexity and performance. Our schemes leverage both stochastic multiplexing gain and electricity-price diversity. Thus, improving over prior work, our schemes reduce energy consumption/cost even when all IDCs have the same electricity price. We use both simulated load traces and real traffic traces to evaluate the performance of the proposed schemes. Results show that our proposed schemes are efficient in reducing energy cost, and robust in QoS provisioning.

Cost Efficient Resource Management Framework for Hybrid Job Scheduling Under Geo Distributed Data Centers

“…Among them, the so-called "capacity right-sizing" is a promising direction, e.g., in [11]. The key idea is to provision servers dynamically based on the load of requests.…”

Energy-efficient virtual machine placement in data centers with heterogeneous requirements

“…where Pid1e is the power consumed by an idle server, Pbus y is the power consumed by a fully loaded server, and E is a constant that depends on the type of physical server [4], [13].…”

“…Furthermore, the CPU utilization levels for thousands of servers was found to be between 10 and 50 percent of their maximum utilization most of the time. Thence, much research has been geared toward reducing the cost of physical servers; for example, Dan et al [4] propose to minimize the energy consumed by first determining the active servers based on the users workload characteristics in a data center, then shifting workload to other data centers if the demands cannot be satisfied. The main idea argued by [5]- [7] is to pack VMs together when it is possible to be able to shut down as many unnecessary servers as possible.…”

Energy-Efficient Virtual Machines Scheduling in Multi-Tenant Data Centers

“…While the energy consumed by a server typically increases linearly with its CPU load, the utilization in Google's data centers for instance has been reported to be only 10 to 50 percent of the maximum CPU load most of the time [4]. Yet, it is well known that even energy efficient idle running servers consume typically more than 50% of the energy they consumes when fully loaded [4], [5]. Therefore distributing the tenants' workloads on the servers intelligently to be able to reduce energy consumption has been a popular approach [5], [6], [7], [8], [9], [10], [11].…”