Just about everything we do nowadays has an online component, and all that virtual reality requires computing power . . . lots and lots of computing power. With the energy demand and costs associated with keeping all those servers running smoothly, data centers are often the earliest adopters of energy-efficient technologies and procedures. And as data centers move beyond the cloud and its basic data collecting functions, a new era of supercomputing awaits. Last month at the Santa Barbara Summit on Energy Efficiency, three industry experts led a panel discussion on data center energy usage entitled, “Energy Efficient Supercomputing and Datacenters at Large Scale.” The panelists offered a wide range of perspectives and prognostications—below some highlights.
In a presentation entitled, “Exascale Computing: Applications Performance and Energy Efficiency,” Horst Simon (Deputy Director, the Lawrence Berkeley National Laboratory) laid out some of the emerging trends in electrical efficiency over time in the area of computer networks. The good news is, “computers have become more energy efficient,” said Simon, pointing to calculations by Stanford University’s Jonathan Koomey that indicate computations per kilowatt-hour have improved by a factor of about 1.5 per year.
But with greater processing power comes greater energy requirements. So, even super computers become more efficient and more power over time—to the point of power/system increases of about 1.24 per year—the power requirements also increase.
As such, Simon warned that, “We won't reach exaflops with the current approach just using standard technology, and the power costs will be staggering to even try.”
Nevertheless, within the sphere of super computing, it is exactly the power requirements of data integration on a massive scale that will drive innovation.
"Industry has not yet pick up on this because they see supercomputer as esoteric," said Simon, “but supercomputer will drive the future of computing.
“There needs to be a mindset change,” he warns, “because supercomputing right now is all about performance, not energy efficiency.”
Pradeep Dubey (Senior Principal Engineer at Intel) elaborated on Simon’s summary of the challenges and opportunities of exascale computing by highlighting some of the projects and research efforts currently being launched by Intel.
Some recent catalysts for change, according to Dubey, include the changing demographics of computer use and the demands for increased computer power. Dubey started out summarizing some of the traditional drivers of computing:
* Norman's Gulf: quest for natural human-machine interface
* Entertainment: unending fascination with virtual and unreal
* The data deluge: the problem of drinking out of fire hydrant
* Real-time analytics: Decision delayed is objective denoted.
* Curious minds want to know (HPC): Science moves on.
Another intriguing aspect of super computing involves the ability to see semantic Web mining (i.e., Google searches and the like) played out in real time. Dubey explained that super computing allows for fundamental shifts in how we uncover and use data. For example, with visual computing, users can access private data, sensory inputs, streams/feeds, immersive 3D graphics output, interactive visualization. One step up, massive data analytics—utilizing a “cloud of servers”—allows for the Intersection of massive data with massive compute real-time analytics, leading to “massive data mining.” And with that massive data and “ubiquitous connectivity,” users can utilize data driven models that are no longer limited to analytical models, but instead are able to access new algorithmic opportunities.
Additionally, Dubey believes that the ability to enhance and elaborate on Web connectivity will further enhance data mining. In particular, “real-time connectivity” enables continuous model refinement, so that even a poor model or can—through massive data and analytic capabilities—be used as an acceptable starting point, since classification accuracy improves over time.
But just as Simon pointed out, Dubey warned that in order to handle the massive data mining possible via exascale computing, traditional power-saving schemes will have to be revisited. This includes increasing “locality awareness,” restructuring in order to improve storage capabilities, and exploring parallel and distributed decomposition in order to improve computer efficiency.
When it comes to massive data mining, Google is the trailblazer and the trend maker. Bikash Koley (Senior Network Architect at Google), began his presentation laying out some basic Google facts:
* Google services are worldwide—over 55 countries and 112 languages.
* Every few hours, Google crawls and refreshes more data than what is contained in the entire Library of Congress.
* YouTube gains 24 hours of video every minute and registers more than 1 billion views a day
“If the Internet were a country,” summarized Koley, “[Google] would be the fifth largest consumer of power in the world—right behind Japan and right ahead of Russia.”
According to Koley, super computing can be house in two types of locations—the traditional data center and the warehouse. In a traditional data center, “infrastructure is decoupled from the structure,” so that there are separate hardware and software systems utilizing “partitioned resources.” In contrast, warehouse-scale computers consolidate computing, so that many user interfaces and applications are housed in one locale. By combining all the hardware and software facets in one system, warehouse-scale computer can utilize a variety of efficient onsite power systems, including renewable energy.
While traditional data centers allow for the storage of massive amounts of data, Koley explained that warehouse-scale computer are, “designed to run massive internet services, applications: tens of binaries running on hundreds of thousands of machines, homogeneous hardware, and system software all utilizing a common pool of resources managed centrally via an integrated design of facility and computing machinery.”
Warehouse-scale computing allows for even greater date center energy efficiency, as exemplified by Google’s own warehouse facilities. While the industry average PUE (total facility power/IT equipment power) is 1.9 and EPA’s target for energy-efficient data centers is 1.2, Google has averaged less than 1.16 for the last 12 months. This increased efficiency was achieved because warehouse-scale computing allowed a few tweaks and retrofits. For example, while cooling traditionally accounting for 30–70% of overhead, Google uses evaporative cooling as a “free cooling” option, thereby avoiding the attendant power requirements of chiller usage. Google also saves by recycling water at its warehouse-scale facilities.
Koley wrapped up by saying that while Google will always aspire to ever-increasing computing capabilities, the company is also aware that proportionate costs for energy will continue to grow. As such, he explained that Moore’s law will most likely conspire with energy demands to keep computing costs roughly fixed. In order to grow, every increase in computer performance and capability will have to be matched by a corresponding increase in the energy efficiency of those systems.
Presentations and video of the Santa Barbara Summit on Energy Efficiency are now available online at http://iee.ucsb.edu/sbsee2011.