Technology with the Environment in Mind

Dynamic Data Center Power Management: Trends, Issues, and Solutions

SOLUTIONS: A POLICY-DRIVEN APPROACH TO POWER MANAGEMENT

A typical data center is managed by deploying management systems for monitoring the health of the computing infrastructure and performing various management functions including error detection, failure resolution, server provisioning, and service deployment to servers. We can extend this management infrastructure to enable the dynamic power management approach described above by introducing a new component: a server-resident PM that meets the requirements for power monitoring and control (see Figure 4).

The PM is responsible for enforcing power management directives that can be specified for each server. This is done by monitoring the appropriate power and thermal sensors on the servers, and by controlling the appropriate effectors that allow the PM to control the servers' power consumption as directed.

Figure 4: Power Management architecture
click image for larger view
 

Intel's current series of server processors contains mechanisms that help control the power consumed by the processors. In addition, memory controllers provide ways of controlling power consumed by the memory modules. These can be used by the PM as low-level effectors for power control.

The interaction model between the PM and the data center management software is shown in Figure 4. The PM receives a policy directive from the data center Management System software. The policy directive may specify a power limit that is to be maintained for the server.

The PM has three major components operating in a closed loop manner: monitoring, policy engine, and control. The monitoring module is responsible for monitoring the sensors to determine if a new set of settings of the effectors needs to be enforced. This is done primarily by comparing the actual values of power and thermal against the policy limits. The policy engine decides a new set of controls based on the deviation of the actual power consumed from the limit and the utilization of various components. The control module then enforces the new settings as determined by the policy engine using mechanisms appropriate for that particular effector.

The sensors used by the monitoring module are the platform power sensors (including system and component powers) and the thermal sensors (system inlet/outlet temperature and component temperatures). Sensors providing information about the utilization of the components are used to determine the behavior of the workload.

The control module uses the effectors described earlier to effect power control on processor and memory. There are also additional effectors for components provided by the OEM (like power supply, fans, etc.) that are used by the PM to achieve greater power control.

When the power consumption exceeds the power limit, the PM will choose a set of settings for various effectors (processor, memory controller, etc.) that effectively reduces the power consumption of the platform. When power consumption goes below the limit, any restrictive controls previously placed on the processor, memory, and other effectors are relaxed. This is done in an iterative manner to account for the fact that the choice of effector settings may not have been accurate given the uncertainties in the behavior of the workload and the uncertainties in the parameters used by the PM to determine these settings. If the PM exhausts the use of all the controls at its disposal and still finds power consumption above the specified limit, it can generate alerts.

The PM exposes an abstracted interface for interacting with the external management software. It also requires interfaces for communicating with monitored components (e.g., temperature sensors, platform power measurements) and controlled subsystems (e.g., processor and memory modules).

The PM can be deployed in various ways:

• As a firmware running on a dedicated microcontroller in the server.
• As part of the baseboard management controller (BMC) that also performs other system management functions in the server.
• As an in-band agent in the operating system.
• As a combination of the above.

While the PM provides the capabilities described in the previous paragraphs, it depends on external management software to specify the policy parameters for it to operate effectively. The PM exposes an interface to the management software for this purpose. The interface includes commands to read power consumption and thermal data. The interface also allows management software to specify commands to set and get power control policies (e.g., set power limit) and to receive alerts from the PM.

The external interface is exposed as extensions to industry-standard server management protocols such as Intelligent Platform Management Interface (IPMI) [14] and Web Services Management (WS-Man) [15]. While IPMI is widely used in the industry today for server management, we anticipate WS-Man to gain more acceptance in the future.

Back to Top