SYSTAP Mapgraph for accelerated Graph Processing

SYSTAP Mapgraph for accelerated Graph Processing : In today’s always on and always connected world, the amount of data that people generate every second is huge and is growing at an incredible rate. Graphs are a way to organize this information by linking it together. Not just nodes and links, and beyond Facebook, LinkedIn, and six-degrees of Kevin Bacon, graphs occur in many important domains and markets. A social network of people and their connections, a fraud-detection system, a battlefield filled with soldiers and other resources, or a human body made of many biological systems are all examples of graphs. The potential of graph analytics is well known, but existing platforms have been limited in their ability to process this information at a large scale. The next generation of significant business, medical, and information science advances will made by those who can harness the data and the links between it quickly and cost-effectively, drawing new conclusions and making discoveries at lightning speed. SYSTAP has a new approach to processing this data, and it’s already showing impressive results. They are delivering graph analytics that process data up to 10,000 times faster using Mapgraph technology.

“Facebook is currently more than a trillion edges and we believe it takes more than 10 minutes using 200+ racks of servers to traverse their graph over a single iteration. We believe our Mapgraph technology could do it in seconds on a cluster of GPUs with similar RAM,” says Brad Bebee, CEO, SYSTAP.

Mapgraph Accelerator is a GPU-based plug-in to SYSTAP’s, open-source graph database platform, Blazegraph. Mapgraph Accelerator marries the speed of GPUs with familiar Java APIs and standardized query languages. It will provide the world’s first and best platform for building graph applications with GPU-acceleration. It will bridge the gap between the Blazegraph database platform and the GPU acceleration for graph analytics. Users of the Blazegraph platform will be able to leverage GPU-accelerated graph analytics via a Java Native Interface (JNI) and via predicates in SPARQL query. Mapgraph HPC is the application of the technology to GPU clusters where it can traverse 100 billion edge graphs in sub-second times.

MapGraph relies on three approaches working together, first, it uses parallel processing. Instead of six or eighteen computer cores processing data at one time, it uses 32,000 cores within each GPU and multiple GPUs. Second, the data is partitioned. Instead of processing every piece of data every time, the data is chunked out like a graph, so the system knows not to process the same data repeatedly. Finally, MapGraph scales to large clusters of GPUs for huge problems. It optimizes the communications between the GPUs using “overlapping communication” patterns. If the communication was not optimized, communicating across all those cores would consume much of the processing power of the system.