The International Supercomputing Conference (ISC’14) has invited one of Japan’s leading HPC experts, Professor Satoshi Matsuoka to deliver a keynote titled “If You Can’t Beat Them, Lead Them – Convergence of Supercomputing and Next Generation ‘Extreme’ Big Data,”
In this thought-provoking talk on Tuesday, June 24, Matsuoka will share why he believes that supercomputer architectures will converge with those of big data and serve a crucial technological role for the industry. His assertion will be exemplified with a number of recent Japanese research projects in this area, including the JST-CREST “Extreme Big Data” project. To understand more about these projects and where they fit into the larger scope of extreme scale computing, we spoke with Matsuoka.
Is there a distinction between “data” and “big data?”
Satoshi Matsuoka: Of course. In fact, I categorize “simple data”, “big data” and “extreme big data” as three different domains.
“Big data” implies two principle characteristics. One is from semantic perspective, in that large data sets are collected in a rather unbiased fashion; and then one would try to extract some meaningful correlative information out of them, using various methods such as data mining, deep learning, graph analytics, etc. Another is from a system perspective, in that the data volume, bandwidth, etc., are too large to be processed with conventional machines, even those geared for traditional databases. The system components, both hardware and software, need enhancements in order to support the increased level of processing. In this sense, big data’s “super data processing” is to normal data processing as supercomputing is to normal computing.
By extreme big data we mean that the data volumes, as well as the computational needs, become so big that a simple extension of conventional big data processing architectures would no longer be feasible and will require convergence with supercomputing technologies and platforms.
How is big data relevant to the HPC space and how has the term evolved over time? Is it something different than what used to be called “data-intensive computing?”
Matsuoka: In some sense HPC has been the pioneer of big data from the days of data-intensive computing. Even as far back as 20 years ago, researchers running climate codes were starting to struggle with terabytes of data when the general public was still in the gigabyte days.
By all means, the general area now covered by big data is much wider. Also due to the emergence of new application areas such as genomics, data-intensive computing in HPC has broadened to the extent that its own coverage has expanded.
How do you envision the convergence between big data and HPC to happen?
Matsuoka: What is unique in the current big data trend is the stress on various data analytics algorithm, such as deep learning and graph analytics. This, coupled with various other factors are requiring some changes to the HPC hardware and software stack, such as the need for a massive increase in data capacity and bandwidth. By contrast traditional HPC is trending toward high bandwidth but low memory capacity.
But since HPC also suffers from lack of memory capacity, the convergence at the hardware level will mostly be in the area of designing capacity-friendly deep memory hierarchies. This applies both to memory depth within a node, using new memory technologies and associated processor architectures, as well as memory width across nodes, requiring extensive use of optics to support high bandwidth and low latency.
From the data side, the needs will be driven by the so-called “broken silos.” Data stored across multiple institutions and disciplines, as well as the proliferation of the internet of things, will cause the data capacities and the compute from the cross-correlations to simply explode. We now have big data applications in genomics that run on almost the entire K-computer, using the abundance of its one-petabyte memory and 660,000 cores. That is already about 1/5 to 1/7 the entire capacity of Amazon according to a major IT consulting company’s estimate. Think of the exascale era when we will have big data apps that require 100 million cores, a number that makes even Google miniscule by comparison.
Right now we have the enterprise with their own application use cases for big data, and perhaps even their own understanding of what the term means. With that in mind, how will a convergence of HPC and big data affect those users?
Matsuoka: Industry also adopts HPC but considers those applications distinct from mainstream computing. By their convergence enterprise and HPC users will learn to better exploit the combined technologies and also allow for the emergence of new applications that tie massive compute to data analytics. We already see examples now in areas such as genomics and design engineering.
Can you please elaborate on Japan’s role in advancing big data technologies and driving its convergence with HPC?
Matsuoka: For Japan, both HPC and big data are high on the agenda for research as well as the industry. It is prudent that we work with other regions of the world with similar vision to push both envelopes. The proposed HPC projects in Japan towards exascale will likely have increased emphasis on extreme big data as well.
Now in its 29th year, ISC is the world’s oldest and Europe’s most important conference and networking event for the HPC community, offering a strong five-day technical program focusing on HPC technological development, and its application in scientific fields as well as its adoption in an industrial environment.
Over 300 hand-picked expert speakers and 170 exhibitors, consisting of leading research centers and vendors, will greet this year’s attendees to ISC. A number of events complement the technical program including Tutorials, the TOP500 Announcement, Research Paper Sessions, Birds of a Feather (BoF) Sessions, the Research Poster Session, Exhibitor Forums, and Workshops. For more, visit www.isc14.org.