Power Down
It's no secret that power consumption is a worrying issue among datacenter
managers. As system hardware becomes cheaper and energy costs continue to rise,
IT managers might find that they'll spend more to power and cool a system over
its lifetime than to actually buy it.
Which is why guys like Dan Pritchett, a technical fellow at eBay, has moved
beyond thinking about transactions per second (TPS) with his applications to
focusing on transactions per second per watt (TPS/w).
"One of the primary challenges we started to face in 2006 was power. The
datacenters were maxed out and we were still running at capacity," said
Pritchett, who noted that local municipalities were often physically incapable
of delivering enough power to meet eBay's growing energy needs.
Power consumption and efficiency are hardly new; mobile platforms have obsessed
about these issues for years. But the need to reduce power consumption in server
and even client applications has become acute enough that some development organizations
are actively seeking ways to do more with less -- be it in the datacenter or
on the desktop.
Check out this
interesting article, titled "The Case for Energy-Proportional Computing,"
for more on the energy consuming habits of modern servers.
Virtualization has played a huge role in datacenter operations, enabling companies
like Google and eBay to maintain ample hardware redundancy while driving up
utilization -- a key for energy-efficient design.
But dev shops can do more, Pritchett said. He urged dev managers to look to
established best practices and to focus on efficient, scalable designs. At eBay,
for instance, databases are always sharded -- split into smaller pieces -- to
produce optimal scaled performance. "Those things definitely come at a cost," Pritchett said.
"But if you are wanting to move into hundreds of millions of entities in
your system, and you're wanting to deal with tens or hundreds of millions of
transactions per day, that's what you are going to move to."
He also urged developers to work toward parallel programming, so that fully
threaded code can work efficiently across multicore processors. "I think
going forward this is definitely going to be a huge issue. We are going to start
having to leverage the parallelization of the hardware into the app space,"
he said.
Are you worried about the energy efficiency of your apps? E-mail me at [email protected].
Posted by Michael Desmond on 07/29/2008