Managing Power in Data Centers

The first thing  to start planning for optimizing the power consumption in

your datacenter is finding out how much power do your datacenter, racks,

servers, equipment, and other devices  consume. Next thing is to figure out the

pattern of the consumption. For instance, what is the time of the day when the

load from a particular rack is maximum or what's the time the overall load of

the datacenter gets minimum. If we  gather such relevant data, then it becomes

very easy for us to plan for better optimization of resources and hence better

optimization of power.

Gathering information



To collect, preserve and analyze information on aspects as mentioned above

you essentially need two components.  First is the power meter which should be

connected to, if not all then at least most of the devices in your datacenter.

Second is the software which can get the feed from these meters and log and

analyze the data.

One simpler way of doing this is by using remotely manageable power outlets

which comes with inbuilt digital power meters which can measure power in terms

of watt and ampere for all the outlets. Such devices come in different form

factors, such as 1U rack mountable chassis, regular power  strips, power strips

which can be installed in the racks, etc.

The added benefit which you get by using such devices is that you can

remotely turn the power supplies on and off over LAN or Internet.  This feature

can help in many ways. Remember, when you had to cold boot a hung server

remotely and you didn't have anyone physically available at the datacenter. This

is a very common scenario which might have happened with many of us.  There are

many companies who are offering such products. These products are called the

metered PDU (Power Distribution Units). One such product which we had reviewed

some time back was  Raritan Dominion PX. This was a 1U box with 8 power outlets

and had a design resembling to a 24 port network switch.

You can even get such PDUs from APC, PDUDirect, WTi, etc. Most of them are

similar in functionality. They generally provide a Web based interface from

where you can cycle power of individual power outlets and at the same time can

see the real time load (both wattage and amperage). Some of these devices do

support logging of historical data locally or remotely but some don't. For those

which don't  have a centralized application which can pool data from various

such PDUs and log them.  One such application is available from Raritan. This is

called PowerIQ. This application not only logs the data from the PDUs but also

tells you active power level, energy consumption, customer costs, environmental

readings, line capacities and carbon footprint. Plus trending and cumulative

total reports are also available at any level including building, floor, room,

rack, user and IT device.

The other benefit this application gives you is that it lets you create

groups of different devices and servers in your datacenter. You can set boot up

and shutdown time for the groups. So for instance if you don't require a bunch

of servers and devices on a particular day of the week or on a particular hour

of the day, then you can turn them off automatically. These devices

automatically will automatically boot up and go live on the scheduled time. This

feature by itself can save a good amount of power.

Redoing the datacenter



The second obvious step towards streamlining business and preventing

leakages in budgets  is  to ensure that the IT systems work efficiently by

incorporating technologies like virtualization, consolidation, migrating to

blade servers and having intelligent energy management solutions on place. In

spite of these deployments, most companies end up saving only a marginal

proportion of the expected outcome, bringing them back to level zero or worse

still, in a few cases, spending more and getting no results. On doing a

ground-up survey, energy efficiency pops up as the biggest concern area, and in

most cases the reason for this is that the IT functions properly, but the

auxiliary 'facilities' like UPS systems, cooling devices and back up power

systems end up unchanged, and result in huge leakages in spite of IT

enhancements.

As an illustration, consider a scenario where 10 servers, that were earlier

running at 50% efficiency percentages, were consolidated to 6 servers running at

80% efficiency each, eliminating the need for 4 servers and consequently the

space and cooling. Even taking into consideration the little extra power a blade

server consumes (assuming the consolidation happened from racks to blade), there

are still tremendous benefits. But if the cooling systems remain unchanged, or

are modified in an incorrect manner, the new architecture will not get

sufficient cooling, resulting in malfunction, or worse, the magnitude of cooling

and power requirements will be disproportionate to the new architecture.

An ideal starting point towards solving this problem is to understand the

magnitude of savings that can be achieved with a mix and match of 'facility'

changes. There many online and free applications which can be found on Google

and which let you indicate the data center capacity, current efficiency, current

UPS systems in use, power costs, lighting details etc, and then allow you to

make indicative changes to any of these and see the savings meter change

dynamically. For example, a 300 KW data center having 50% IT load, at typical

Indian electricity charges, running on legacy UPS systems, with chilled water

cooling system at single path power is likely to run at an efficiency of 43.8%

with an annual electricity cost of $353,050. If row-based cooling is deployed

and a high efficiency cooling system along with blanking panels and

'intelligent' lighting in the server room is installed, the savings percentage

climbs up to 58.6%, bringing down the annual energy cost to $263,639. Whether

the IT manager sees sense in these deployments for the overall cost benefit the

company would derive is his/her call, but the tools are reasonably comprehensive

in its analysis of loss reductions.

This is only the starting step in understanding or rather deriving the

problem. Having understood the area where energy efficiency can and should be

incorporated, the next step is to identify relevant technologies that can be

used to achieve better efficiency. One of the simple methods an enterprise can

deploy is the Hot Aisle data-center architecture. In a scenario where an

enterprise spends a major chunk of its data-center budget in raised flooring and

air conditioning for the servers — blades or otherwise — savings can

straightaway be realised if the Hot Aisle arrangement is used. The fans of a

server release heat into the room and the faster the server performs, the more

is the heat released. This heat is the single biggest contributor to the

requirement of a redundant air cooling system, so much that the heat released by

all the server fans put together is directly proportional to the cooling

expenditure of the data-center. Hot Aisle architecture is merely the

arrangements of servers in such a way that two rows of servers are having the

fans facing each other. The alley (or aisle) between the two fans becomes the

Hot Aisle and the next alley (the front panel of the server) becomes the Cold

Aisle. First, this architecture ensures better measurement of heat released and

gives a better idea of cooling requirement for the data center concerned and

second, it collects hot air in a separate corridor, and this trapped air can be

'fed' into the cooling system for better efficiency. The cold aisle, meanwhile

only needs re-circulated air to keep the servers functioning. Most

state-of-the-art data centers that deploy this architecture also have an

artificial roof for the hot aisle to trap the more air more efficiently. From a

data-center design point of view, this architecture gives a better indication of

the 'hot spots' or 'red' areas of the data center and allows for better planning

and expansion.

Another technology that CIOs and data center managers can deploy and use is

In-Row cooling. This technology goes on to prove that modifying the cooling

architecture in a data-center can go a long way in achieving better power

efficiency. Traditional cooling mechanisms utilize the raised floor of a

data-center to pump air from the bottom and cool each server in the rack bottom

upwards. With the advent of blade servers, the performance of each other and the

corresponding cooling it requires is directly proportional to the compute power

that is designated to the particular blade. In such a scenario, the blade that

is sitting right on top of the stack may be the one releasing maximum hot air,

as it is the one that performs the most. This creates an artificial 'hot zone'

around it, which is further magnified if hot aisles are absent. A simple way to

solve this problem is to cool sideways — in other words, cool air can be pumped

not bottom upwards but along the sides of each server. This way, each server in

the rack across multiple racks gets the same amount of cooling. This may not

always be needed, so automating In-Row cooling of this nature can help the

cooling system 'understand' which server is performing faster than the other,

and adjust the cooling dynamically in real time. If the task assigned to the

particular server is completed and load is shifted to a different server in a

different rack, the cooling changes accordingly to cool the newly assigned

server more effectively. In data-center scenarios where compute power is

required in spikes — for instance on an share broking site where load reduces

significantly on weekends — In-Row cooling can directly contribute to reducing

the power expenses of the enterprise.