DARPA targets big data with graph processor effort

Chipmaker Intel Corp. and four others will develop a graph analytics processor to help find relationships in big data.

A new processor architecture tightly focused on data-intensive applications is advancing under a DOD effort that seeks to wrangle the exploding number of data types ranging from sensor information to video files.

The Defense Advanced Research Projects Agency unveiled the "graph analytics processor" initiative last summer and recently selected five contractors to begin development under a program called Hierarchical Identity Verify Exploit, or HIVE.

HIVE developers include chip giant Intel Corp. and Qualcomm Intelligent Solutions, integrator Northrop Grumman along with Pacific Northwest National Laboratory and Georgia Tech. Researchers at the Atlanta-based university have done early theoretical work on graph computing as a "crucial tool for processing big connected data."

The sheer volume of data and emerging data types are overwhelming current processor architectures widely based on Intel's dominant x86 architecture and graphics processors, or GPUs, from chipmakers like Nvidia Corp. So, DARPA is looking for more efficient ways to churn through the torrent of sensor and other so-called unstructured data.

The goal of the HIVE program, according to the DARPA program manager, Trung Tran, is developing a new data-handling platform for analyzing and making sense of huge data volumes. "Current single-chip CPU [and] GPU hardware cannot efficiently process large graphs in real time," Tran stressed. Overcoming current processor limitations requires large datacenters.

HIVE also seeks to move beyond the current hardware focus on "dense data" accessed sequentially by chip memory. In order to leverage what Tran called "sparse data" such as graphical models and decision trees, developers will adopt a random access approach to boost chip performance.

Graph processing is an emerging data analytics technology used for applications like cyber defense and critical infrastructure protection that require analyzing huge data sets in real time. The commercial market for graph databases has exploded over the last several years as data scientists leverage inherent performance advantages to uncover relationships within huge data sets.

Among the ways of querying all that data are approaches that discover how snippets of data may be connected. Graph analytics excels at detecting relationships within data. In a recent survey released by IBM, commercial users cited speed and performance as the top graph technology attributes. A dedicated graph analytics processor promises to extend those performance gains as traditional processor architectures run out of steam, DARPA officials noted.

"Although efficient implementations of specific graph applications exist, the behavior of full-spectrum graph computing remains unknown," Georgia Tech researchers noted in a paper delivered at the 2015 supercomputing conference. "To understand graph computing, we must consider multiple graph computation types, graph frameworks, data representations, and various data sources…."

Tran, the DARPA program manager, said he expects the processor effort to yield a hefty 1,000-fold improvement in big data processing efficiency. "This will enable relationships between events to be discovered as they unfold in the field rather than relying on forensic analysis in datacenters," he added during a briefing for contract bidders.

The projected HIVE processor would differ from current processors by employing parallel memory access along with standard parallel processing techniques. That approach is intended to move memory and computing resources closer to relevant data to uncover relationships. The upshot, Trans predicted, is "scalable, real-time graph analytics at the network edge."

The DARPA project also could leverage emerging memory architectures such as Intel's 3-D Xpoint non-volatile memory technology rolled out in commercial solid-state drives earlier this year.

Meanwhile, the combination of chip companies, a large DoD supplier and integrator along with university and government researchers is intended to "forge new R&D pathways that can deliver unprecedented levels of hardware specialization," noted William Chappell, director of DARPA's Microsystems Technology Office.

The mix of commercial and military players also is intended to strengthen the military electronics supply chain, Chappell added, as current hardware and software development struggles to keep pace with the rise of big data.

The concept phase of the HIVE program extends though next year, with initial prototyping beginning in fiscal 2019. Chip fabrication could begin as early as fiscal 2020, the agency said.