Power Management in Servers

Lately one of the major challenges enterprises are facing is to bring down

cost of power consumption and cooling in their data centers. Even determining

how much power a particular server in a data center is consuming is very

difficult. In a typical scenario, a fixed amount of power is supplied to the

room and from there onto the racks. The challenge here is to determine how many

servers can be placed in a single rack without exceeding the limit. To achieve

this, Data center managers typically go by the power consumption value present

on the server. However, servers usually consume lesser power than what is

written on their nameplate, and this scenario also leaves racks under utilized.

Solution to this is simple. Data center managers require more visibility in to

how much power is being consumed by the servers. They need to constantly monitor

power consumption to better understand this server/rack utilization and perform

power budgeting.

In a server, power management can be divided in various ways. One common way

to divide this is at processor level, Chipset level, Operating System level and

application level. To solve the aforesaid issues, there is a lot of development

going on in reducing power consumption of servers at each level, and dynamic

power management of servers. Vendors are looking for ways to lower the power

consumption of servers according to the workloads being processed by them and

dynamically reallocate power to servers when their workloads need more

processing power. Here we will be looking at various new technologies related to

power management of servers that have hit or are about to hit the market, which

aim to improve upon power consumption considerably.

Power management in Intel's Nehalem platform



Nehalem has an integrated micro-controller called Power control Unit or PCU.

It has its own embedded firmware and dynamic sensors to monitor current,

temperature and power in real time. Nehalem also has an integrated power gate

which eliminates the problem of power leakage. Power enables per core C6 state

i.e. every individual core can switch to near zero power state. This process is

invisible to the OS, software as well as other cores. This functionality is

further extended with Nehalem's turbo mode feature in which when idle cores are

shut, power is channeled to active cores in order to boost their performance.

Source: Intel Nehalem has an integral power control unit with its own embedded firmware and dynamic sensors to monitor power in real time.

Intel Dynamic Power Node Manager



This is an out-of-band power management policy engine which is embedded in

Intel server chipset. It works along with the BIOS, OS and power management

functions. It measures server's actual power consumption with a margin of plus

or minus 10% and generates reports. Users can track how much power is consumed

by a server over a particular period of time. Let's say a server consumes a

maximum of 300 watts over a period of time and the budgeted power for the server

is 600 watts, knowing this you can easily lower budgeted power to 400 watts.

Node manager also supports power capping. It can enforce a policy set in server

management console. It receives the policy through IPMI interface and maintains

power by adjusting the CPU p-states.

AMD Smart Fetch



Bit similar to Intel, AMD Shanghai has a technology called Smart Fetch,

which can power down idle cores. As per AMD, this will reduce Power consumption

of CPU by 21%. It also has AMD Cool core technology that reduces the power

unused part of the processor, to reduce power consumption further. Next version

of AMD server processor Istanbul might have a feature called Cool Speed. This

processor provides protection to processor integrity by reducing the power

states when it reaches a temperature limit. This will help platform providers in

reducing system fan speeds which eventually could result in better efficiency.

Example of P-state being used to reduce power consumption in Windows Server 2008 R2.

Core parking



This is a new development in Windows server 2008 R2, which is currently

available in its Beta. Core Parking constantly tracks workloads processed by

every core, and consolidates the workload to a minimum possible cores and

deactivates the inactive cores, thus saving on the un-necessary power being

consumed by the inactive cores. Similarly when suddenly a workload requires more

processing power, it activates the remaining cores. Along with this, Beta of

Windows server 2008 R2 can reduce server power consumption by modifying ACPI

'p-states'.

Also with this, power consumption of the server can be monitored through the

system center and power budgeting can also be performed. Just to refresh your

memory, P-States' are performance states defined under ACPI specification. More

details about ACPI can be found at www.acpi.info