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

Abstract

Despite the dramatic improvements achieved in building energy-efficient electronic devices, the amount of electricity consumed worldwide to power the global information technology infrastructure has grown tremendously in the past decade. In this paper, we propose algorithms to reduce energy consumption by data centers by considering the placement of virtual machines onto the servers in the data center intelligently. We formulate this problem as an integer programming problem, prove it is NP-hard, then explore two greedy approximation algorithms, minimum energy virtual machine (VM) scheduling algorithm (MinES) and minimum communication virtual machine scheduling algorithm (MinCS), to reduce the energy while satisfying the tenants' service level agreements. We examine the performance of these two algorithms in both small and large clusters using real data traces and synthetic workloads, and compare them to other alternatives. Our results demonstrate that MinES and MinCS yield scheduling that are within 4.3% to 6.1% energy consumption of the optimal solution while being computationally efficient.

Energy efficient temporal load aware resource allocation in cloud computing datacenters

“…The assignment of VMs to the servers and flows to the links in clusters are performed using a heuristic. The [11] also studied the joint optimization of server and network power consumption. They formulated the problem as an integer programming problem, proved that it is NP-hard and then proposed two greedy algorithms for VM scheduling.…”

The influence of datacenter usage on symmetry in datacenter network design

“…In Oktopus, the focus is on providing bandwidth guarantees for virtual datacenters in the form of virtual machines all connected to a virtual switch, called virtual clusters in [4], and also collections of virtual clusters where the virtual switches of these clusters are connected to a root virtual switch, called oversubscribed virtual clusters in [4] (as is noted in, for example, [14], the Amazon EC2 cloud embeds tenant requests in the form of a virtual cluster). It is remarked in [4] that other virtual datacenters in the (topological) form of hypercubes, multidimensional meshes, de Bruijn networks, and so on might also be offered to users (though this possibility is not seriously examined).…”

“…ant colony optimization algorithms in [66], greedy algorithms in [48], and linear programming in [14,65]. Of course, with a fixed DCN as the host, it is possible that the virtualization problem becomes solvable in polynomial time (though, in our view, unlikely for existing DCNs).…”

“…As we stated earlier, as yet there have been no attempts to tackle virtualization in a server-centric setting; however, the energy models as regards virtualization and energy efficiency in switch-centric DCNs hold good in server-centric DCNs. The fundamentals concerning energy efficiency are well laid out in [14], for example. Broadly speaking: the energy consumption of a server depends upon the CPU load, with the idle server still consuming a significant fraction of the energy consumed when it is fully loaded; and the energy consumption of a switch is dependent upon the number of ports that are disabled, with a switch with all its ports disabled still consuming a significant fraction of the energy consumed when all ports are enabled.…”

“…Broadly speaking: the energy consumption of a server depends upon the CPU load, with the idle server still consuming a significant fraction of the energy consumed when it is fully loaded; and the energy consumption of a switch is dependent upon the number of ports that are disabled, with a switch with all its ports disabled still consuming a significant fraction of the energy consumed when all ports are enabled. The upshot, from [14], is that (not surprisingly) it makes sense to run CPUs with as high a load as possible and to enable as many switch ports as possible. Note that our methodology supports both intuitive aims: in an unfragmented embedding of virtual clusters along our path, there is at most one switch-node with an enabled port and a disabled port, and there is most one server-node where the CPU is not idle and not fully loaded.…”