Cisco Cisco UCS C22 M3 Rack Server White Paper

Page 4 of 15

Three major variables affect the optimization of power usage in a data center

●

Cooling and distribution system efficiency: The cost of cooling the system is a sizable portion of the cost

of the energy used to perform the workload.

●

Efficiency: Although the overall power utilization efficiency (PUE) of a data center is outside the scope of

the server design, the server design should still try to achieve maximum compute efficiency per watt of

power used.

●

Capacity utilization: It is important to have an accurate measure of the power provisioned for a system.

Under-provisioning and overprovisioning can have a negative influence on data center performance. While

the former can severely affect critical business operations at peak hours, the latter can lead to massive

underutilization of infrastructure.

Apart from the operational efficiency aspects of power management, it is important to keep in mind the data center

design aspect, such as over-current circuit protection. While compute end users and IT administrators may be

concerned with the amount of power consumed per server and per chassis, the data center operations people are

more sensitive to total power per circuit and the ability to protect circuit breakers from tripping.

A great deal of literature exists on the subj

ect of power provisioning in data centers, including “Power Management

in the Cisco Unified Computing System: An Integrated Approach”

and “Data Center Power and Cooling”

. A data

center designer has to make multiple assumptions about the nature of the data center workload when provisioning

the servers across the power buses.

Administrators and designers may assume that not all servers will draw the maximum power at the same time and

also that, even at 100 percent utilization, they will not all draw the maximum power for which the system power

supplies are rated. These assumptions help the designer oversubscribe the available power with respect to the

number of systems deployed. Unfortunately, the nature of workloads in data centers is changing rapidly with

widespread deployment of technologies such as virtualization and cloud-based storage and computing. These

technologies are attempting to drive down the total cost of ownership (TCO) while maximizing the utilization of

hardware infrastructure. The new workloads are making it more difficult to sustain the assumptions that allowed

designers to oversubscribe their power supply infrastructure. Modern workloads are displaying increasingly

unpredictable fluctuations in traffic and power load spikes

Power Capping in Cisco UCS C-Series M4 Servers

Designers need to have easy access to key metrics of infrastructure utilization, in a convenient and reusable form.

These metrics include power consumption patterns over periods of days; the maximum, minimum, and average

trends for the entire platform; the distribution of consumption across subsystems in the servers (Including CPU,

memory, storage, and I/O); and the actual possible maximum power consumption of the server configuration. In

rack-mount servers, which are open configuration systems with a huge matrix of possible on-board peripherals,

attempting to get these numbers right has been a difficult and error-prone exercise. The penalty for miscalculating

is severe

—underprovisioning the power per system per rack might lead to a circuit breaker trip, eventually causing

a data center outage. Overprovisioning, on the other hand, drives down the PUE of a data center. The Cisco UCS

C-Series M4 servers attempt to help designers solve these problems by providing more accurate and precise

telemetry and statistics.

The technology used to gather this information is built into the hardware of the systems, allowing levels of precision

not possible in legacy technologies. Additionally, this solution is capable of scaling at the system level across

different configurations, and at the data center level by providing an XML API for management and reporting.