Big data and big compute are not new concepts. Before the term “big data” took off as the buzzword du jour, the HPC community expressed these same ideas as compute-intensive and data-intensive computing. Problems were compute-bound or IO-bound or both.
It is the case, however, that the world is in the midst of a data explosion. In 2013, the amount of data flowing through the Internet was 667 exabytes, an amount equivalent to more than 141 billion DVDs. The quick rise of the big data conceptual framework reflects this paradigm. Big compute works nicely as a complementary term. They are essentially two sides of a coin, or are they?
In a recent TEDx Talk, Virginia Tech professor and noted HPC expert Wu Feng discusses how these elements are experienced differently across nations.
Feng begins his talk with a question: “In today’s rapidly evolving technological world, is our future in big data or big compute?”
As he provides an overview of the terms, Feng references HokieSpeed, the GPU-accelerated supercomputer that he developed, which debuted as the greenest commodity supercomputer in the US in November 2011. HokieSpeed is a big compute resource, notes Feng, capable of calculating 500 trillion operations per second*, 100,000 times faster than a typical PC.
HokieSpeed and other systems like it are being used for epidemiological studies, which can be used to guide public policy in the event of disease outbreaks. Simulations boost scientists’ understanding of how viruses spread, enabling them to assist public health officials in devising appropriate containment measures.
Another HokieSpeed project aims to reverse-engineer the brain. Researchers are trying to find repeating patterns of higher-order motor function in EEG brain readings. Simulations are used map neurological pathways.
One of the neurological ailments in the news today is called CTE, a progressive, degenerative brain disease that is affecting athletes with a history of brain trauma, namely concussions. CTE can only be definitely diagnosed after death, but neurologists are working towards diagnosing and treating CTE in living patients. On a PC, this kind of research would take months or years instead of hours or days.
Big data has many definitions, and one important characteristic is that it’s relative, i.e., more data than you are used to. “Big data is your humongous haystack and various algorithms that you use to root around that haystack. Big compute is lots of metal detectors,” explains Feng. “They’re the devices with which you are going to try and find all the little needles of information in the haystack that you can glean some insight and knowledge from.”
Feng makes the case that different nations have different priorities when it comes to investing in big data or big compute.
Back in May 2013, Feng spoke with White House officials to discuss DNA sequencing research in the life sciences. One of the applications here includes finding mutations in genomes. This makes it possible to then infer different pathways that are causing cancer, setting the stage for potential treatments. At this function, there was clearly a focus on big data, notes Feng, while big compute, while important, was clearly secondary.
Three weeks later, Feng traveled to China as part of a US delegation, where he found that the converse was true.
“Here, we look at big data as being more important,” Feng states. “And in China, big compute is more important than big data, so much so that they created a supercomputer called TIANHE-2 that is 282 times faster than HokieSpeed and twice as fast as the fastest US supercomputer.”
They view big data merely as an application area of big compute, notes Feng.
Feng contends that big data, at least in the US, has been elevated to a position above big compute, in part because the compute side is so often hidden from the user. For example, Google returns search results with lightening speed, but the average person does not realize the immensity of the underlying computational infrastructure that has enabled this transaction.
He cites IBM Watson’s Jeopardy appearance as another example of a very visible “big data” application where the compute side was essentially hidden from the audience.
So what should we be investing in? asks Feng. As complementary forces, the data and compute go hand-in-hand. “In order to make sense of the data, we need to compute on the data.” There is a cycle in which data becomes information, then knowledge, then wisdom – and each of these steps requires computing.
*Note: According to Virginia Tech’s announcement, HokieSpeed claims “a single-precision peak of 455 teraflops, 455 trillion operations per second, and a double-precision peak of 240 teraflops, or 240 trillion operations per second.”