Communications of the ACM
Catalyst Supercomputer Available for Collaborative Research
Lawrence Livermore's Catalyst supercomputer represents a major departure from classic simulation-based computing architectures common at U.S. Department of Energy laboratories.
Credit: Lawrence Livermore National Laboratory
Catalyst, a first-of-a-kind supercomputer at Lawrence Livermore National Laboratory, is available to industry collaborators to test big data technologies, architectures, and applications.
Developed by a partnership of Cray, Intel, and Lawrence Livermore, this Cray CS300 high performance computing (HPC) cluster is available for collaborative projects with industry through Livermore's High Performance Computing Innovation Center (HPCIC).
"Over the next decade, global data volume is forecasted to reach more than 35 zettabytes," says Fred Streitz, director of the HPCIC. A zettabyte is a trillion gigabytes. "That enormous amount of unstructured data provides an opportunity. But how do we extract value and inform better decisions out of that wealth of raw information?"
A resource for the National Nuclear Security Administration's (NNSA) Advanced Simulation and Computing (ASC) program, the 150 teraflop/s (trillion floating point operations per second) Catalyst cluster has 324 nodes, 7,776 cores, and employs the latest-generation 12-core Intel Xeon E5-2695v2 processors. Catalyst runs the NNSA-funded Tri-lab Open Source Software that provides a common user environment across NNSA Tri-lab clusters at the Los Alamos, Sandia, and Lawrence Livermore national labs.
"The opportunity to work with Cray and Intel to design and deploy Catalyst, a novel computing platform optimized for HPC-end applications, has been very exciting," says Robin Goldstone, Livermore HPC Solutions architect. "We have modified the Cray CS300 architecture in ways that make Catalyst an outstanding HPC platform for data-intensive computing."
Catalyst features include 128 gigabytes of dynamic random access memory (DRAM) per node, 800 GBytes of non-volatile memory (NVRAM) per compute node, 3.2 terabytes of NVRAM per Lustre router node, and improved cluster networking with dual rail Quad Data Rate Intel TrueScale fabrics. The addition of an expanded node local NVRAM storage tier based on PCIe high-bandwidth Intel Solid State Drives allows for the exploration of new approaches to application check-pointing, in-situ visualization, out-of-core algorithms, and big data analytics. NVRAM is familiar to anyone who uses USB sticks or an MP3 player; it is simply memory that is persistent and that remains on files even when the power is off, hence "non-volatile."
Deployed in October 2013, the Catalyst architecture already has begun to provide insights into the kind of technologies the ASC program will require over the next decade to meet high performance simulation and big data computing mission needs. The increased storage capacity of the system (in both volatile and nonvolatile memory) represents the major departure from classic simulation-based computing architectures common at U.S. Department of Energy laboratories and opens new opportunities for exploring the potential of combining floating point focused capability with data analysis in one environment. The machine's expanded DRAM and fast, persistent NVRAM are well suited to a broad range of big data problems including bioinformatics, business analytics, machine learning, and natural language processing.
Jonathan Allen, a Lawrence Livermore bioinformatics scientist, is working on new methods to rapidly detect and characterize pathogenic organisms such as viruses, bacteria, or fungi in a biological sample.
"We're working on developing scalable analysis tools for next generation sequencing, in particular metagenomic sequencing," Allen says. "By comparing short genetic fragments in a query dataset against a large searchable index of genomes, we can make determinations about the potential threat an organism poses to human health."
Traditional technologies and storage limitations made it challenging to rapidly search a database of reference genomes as more organisms were sequenced and more variants in the population of an organism were included. With Catalyst's unique architecture, Allen and his team are able to store very large reference databases of genomes in memory and execute expansive analyses with higher resolution.
"We were able to do a metagenomic analysis on a fairly large sample in several hours on a single desktop. With Catalyst, we can process many hundreds of equal size in about the same time."
Catalyst also will serve to host very large models for video analytics and machine learning.
"YouTube claims that 100 hours of video are uploaded to its website every minute," says Doug Poland, computational engineer working on video analytics. "As the fastest-growing type of content on the Internet, consumer-produced videos are a wealth of information about the world that's essentially untapped."
Yet current tools are unable to search through the richness of video elements such as visual, audio, and motion, and associated metadata like semantic tags and geo-coordinates. Poland and his team are looking to build more complex models that consider the sum of those features, and that can be recognized in real-time for user-specific search needs.
"Catalyst allows us to explore entirely new deep learning architectures that could have a huge impact on video analytics as well as broader application to big data analytics," Poland says.
Streitz adds, "Our purpose is to use Catalyst as a test bed to develop optimization strategies for data-intensive computing. We believe that advancing big data technology is a key to accelerating the innovation that underpins our economic vitality and global competiveness."
Companies interested in access to Catalyst are invited to respond to the posted Notice of Opportunity through the Federal Business Opportunities website and visit the HPC Innovation Center website.
No entries found