MARCO CONCEPTUAL

As mentioned in Section 3.1, minimizing SLA violations and minimizing power consumption are commonly studied goals in data centre management. In this section, we will discuss the design of management strategies to pursue these goals, which are representative of those found in the literature. We design two single-goalmanagement strategies,Power andSLA, to work towards achieving a single goal each. Then, we present one Hybrid strategy, which makes a best effort to achieve both goals simultaneously, still within a single strategy. The strategies presented assume frequent monitoring. Calculations are performed on monitored values over a sliding window of time, referred to as themonitoring window.

3.3.1

Terminology

This section presents the terms and metrics used in the description of management strategies.

SLA Violation

An SLA violation, denoted Sv_{, occurs when resources required by a VM are not available to}

it, as this situation leads to a degradation in performance. The percentage of required CPU not available is the SLA violation, as described in Chapter 2.6.

Data Centre Utilization

The overall utilization of the data centre is calculated as the percentage of total CPU capacity in the data centre that is currently in use.

3.3. ManagementStrategies 27

Power Efficiency

For a host,h, the power efficiency,ψh, is the amount of processing being performed per watt

of power. This is measured in CPU-shares-per-watt (cpu/watt). LetΩ0_h be the CPU utilization of host h, and Ψh be its power consumption. Power efficiency (ψh) of a single host is then

calculated as follows:

ψh =

Ω0

h

Ψh

As an active host machine consumes a significant amount of power even when under little or no CPU load (i.e. very low power efficiency) increased host utilization corresponds with increased power efficiency for that host. This metric is used to calculate the power efficiency for the entire data centre,ψdc, calculated as

ψdc= X h∈H Ω0 h X h∈H Ψh

Maximum Power Efficiency

This metric represents the best power efficiency a host can achieve, calculated as the power efficiency of the host at maximum CPU utilization, and is denotedψmax_h .

Optimal Power Efficiency

Optimal Power Efficiency,ψopt_dc, represents the best possible power efficiency achievable at the data centre level, given the current workload and set of host machines available. The best power efficiency would be achieved by placing VMs in such a way that each host is 100% utilized, with the most power efficient hosts being filled first. We first calculate the total CPU- in-use across the data centre. We order the available hosts by maximum power efficiency, and allocate the CPU-in-use to hosts such that each host is allocated 100% of its CPU capacity. We calculateψopt_dc to be the power efficiency of the data centre given this allocation.

3.3.2

Power and SLA Strategies

ThePowerandSLAstrategies aresingle-goalstrategies, which means that all management de- cisions are geared towards achieving a single, primary goal. Single-goal strategies may pursue secondary goals, but always give them lower priority than the primary goal. The presented strategies are based on the management operations and algorithms presented in Chapter 2.4.

28 Chapter3. SwitchingDataCentreManagementStrategies atRun-time

These strategies have several opportunities to differentiate themselves in order to pursue their individual goals, and the choices are reflected in the individual policies of each strategy. Strate- gies can be tailored to a specific goal through the following methods:

• Setting appropriate threshold values for classifying hosts asstressed(the upper threshold, Ωτ) andunderutilized(the lower threshold,Ωτ).

• Modifying the sorting order of target hosts in the three management operations. • Setting the execution frequency ofVM RelocationandVM Consolidationoperations. See Chapter 2.4 for details on host classification and management operations.

In the following sections, we describe the specific settings for the policies of each strategy, and how they work to achieve the goal of the strategy.

Host Classification

Each strategy uses a different value for the upper threshold, Ωτ, which controls the point at which a host is classified asstressed. Thepowerstrategy setsΩτ = 95%, allowing hosts to be very highly utilized before triggering a stress situation and subsequent VM migration. TheSLA

strategy setΩτ= 85%, which provides an extra safety margin to handle increases in VM CPU demand. Both strategies use a lower threshold value ofΩτ = 60%, which defines the point at which a host is classified asunderutilizedand is a candidate forVM Consolidation.

VM Placement

TheVM Placementoperation runs each time a new VM creation request is received, and selects a host in which to place the new VM. See Algorithm 3 for the general placement algorithm. We modify the construction of the target host list (line 2) for each strategy. ThePowerstrategy sorts the partially utilized (H+) andunderutilized (H−) sets in decreasing order first by maximum power efficiency (ψmax_h ), and then by CPU utilization (Ω0_h). Theempty(H∅) host set is sorted in decreasing order by power efficiency. Finally, the target list is built asH+·H−·H∅. This sorting method ensures that the placement focuses on power efficiency over any other considerations.

The SLAstrategy constructs its target host list differently. The partially utilized hosts are sorted in increasing order first by CPU utilization and then by maximum power efficiency, and theunderutilizedhosts are sorted in decreasing order first by CPU utilization, and then by max- imum power efficiency. The sorting order of theemptyhosts, as well as the final construction of the target list, remains the same as in thePower strategy. This sorting method ensures that the placement focuses on spreading load across the set of hosts, leaving spare resources to handle spikes in resource demand, at the expense of other goals.

3.3. ManagementStrategies 29

VM Relocation

TheVM Relocationoperation responds tostressedhosts by migrating a VM to free additional capacity and eliminate the stress situation. Both strategies execute the operation every 10 minutes. See Algorithm 1 for the general VM relocation algorithm. We modify the host clas- sification (line 1) and target host list construction (line 3) for each strategy, and both strategies sortstressed hosts (line 2) in decreasing order by CPU utilization (Ω0

h). During host classifi-

cation, thePower strategy considers a host to bestressed if its CPU utilization has remained above theΩτthreshold over a specified monitoring window. This helps ensure that a migration is not triggered due to a transient spike in demand, but also may result in a slow reaction to a sustained increase. It constructs the target host list in the same manner as described for theVM Placementoperation.

The SLAstrategy, on the other hand, considers a host to bestressed if its CPU utilization exceeds theΩτ threshold in the last recorded monitoring value, or on average over a specified monitoring window. It constructs the target host list in the same manner as described for the

VM Placementoperation.

VM Consolidation

TheVM Consolidationoperation consolidates VMs onto the fewest number of hosts possible by migrating VMs off of underutilized hosts and switching them into a power saving mode (i.e. off or suspended). This operation executes less frequently then VM Relocation. See Algorithm 2 for the general VM consolidation algorithm. Both strategies construct their target host lists (line 3) in the same manner as in their respectiveVM Placement andVM Relocation

policies. Furthermore, the Power strategy executes VM Consolidation relatively frequently, on a 1 hour interval, in order to aggressively consolidate load. The SLA strategy executes

VM Consolidation only every 4 hours, thereby decreasing the risk of overloading hosts by consolidation.

3.3.3

Hybrid Strategy

We designed a dual-goalstrategy as a combination of the Power and SLA strategies; theHy- bridstrategy consists of theVM PlacementandVM Relocationpolicies of theSLAstrategy and theVM Consolidationpolicy of thePower strategy. Furthermore, the stress check performed by theVM Relocationpolicy represents a compromise between the checks ofSLAandPower: it determines that a host is stressed only if its average CPU utilization over the last monitor- ing window exceeds theΩτ threshold. The thresholds Ωτ and Ωτ were set to 90% and 60%, respectively, as a compromise between the values set for thePower andSLAstrategies.

30 Chapter3. SwitchingDataCentreManagementStrategies atRun-time