Thursday, November 19, 2015

OpenPOWER's Order–of-Magnitude Performance Improvements

By Rich Ptak and Bill Moran

Performance improvements come in different sizes. Often vendors announce a 20% or 30% performance improvement along with an increase in the price/performance of their product or technology. Much more rarely, a vendor delivers an order-of-magnitude improvement. An order-of-magnitude improvement equates to a performance increase of a factor of 10. Improvements on this scale underlie recent[1] technology acceleration announcements[2] by IBM and other OpenPOWER Foundation members.

Why are tenfold performance improvements especially important? Here’s why.  Consider this transportation example of what an order-of-magnitude change means. Let’s say running can be sustained at a rate of 10 miles per hour. An order-of-magnitude change raises that to 100 miles per hour. Many cars can achieve and maintain that speed. (We aren’t recommending that!) Another order of magnitude improvement in speed moves us to a jet airplane at 1,000 miles per hour. Another increase of this magnitude moves to a rocket reaching 10,000 mph.

Notice that each magnitude change increases not just speed, but dramatically transforms a whole landscape. Moving from the jet to the rocket allows escape from earth’s atmosphere to go to the moon. This demonstrates the potential importance of order-of-magnitude improvements. The OpenPOWER announcements detail multiple such improvements, let’s examine a few.

One example comes from Baylor College of Medicine and Rice University announcing breakthrough research in DNA structuring[3]. The discoveries were made possible by an order-of-magnitude improvement in processor performance. As reported by Erez Lieberman Aiden, senior author of the research paper, “the discoveries were possible, in part, because of Rice’s new PowerOmics supercomputer, which allowed his team to analyze more 3-D folding data than was previously possible.” A high-performance computer, an IBM POWER8 system customized with a cluster of NVIDIA graphical processing units “allowed Aiden’s group to run analyses in a few hours that would previously have taken several days or even weeks.”

Another example involves IBM’s Watson and NVIDIA’s Tesla K80 GPU system[4]. Watson[5], of course, is IBM’s leading cognitive computing offering which runs on IBM OpenPOWER servers. NVIDIA’s new system allows Watson’s Retrieval and Rank API to work at 1.7 x its normal speed. Wait a minute you might say. Where is the order-of-magnitude change here? 1.7 is impressive, but it’s no order-of-magnitude change.

Almost as an afterthought, IBM mentions that the GPU acceleration also increases Watson’s processing power to 10x its former maximum. So there we have another tenfold improvement in performance arrived at by marrying other technologies to Power.

Finally, Louisiana State University published a white paper[6] stating that Delta, its OpenPOWER-based supercomputer, accelerates Genomics Analysis by increasing performance over their previous Intel-based servers by 7.5x to 9x. Not quite an order-of-magnitude, but close. 

The announcement includes more examples demonstrating the potential of the OpenPOWER philosophy, OpenPOWER Foundation and Power Systems to achieve dramatic results across multiple industries. The fundamentals of the POWER architecture lead us to anticipate continued improvements in Big Data processing. Such developments will accelerate the growth of the Internet of Things. It will also drive fundamental changes in the possible types of processing, just like those happening with Cognitive Computing.