A plenary panel at the XSEDE15 conference, which took place this week in St. Louis, Mo., highlighted the broad spectrum of computing resources provided by the National Science Foundation, including several new and testbed projects and an effort to help more people use cyberinfrastructure to advance their research.
“I don’t think there has been a time previously when NSF funded the diversity of systems that are available today,” said panelist Craig Stewart, the associate dean of research technologies at Indiana University.
Irene Qualters, leader of the Division of Advanced Cyberinfrastructure within NSF’s Computer & Information Science & Engineering Directorate, kicked off the panel with an overview of how “the conduct and the practice of research are changing,” and how this is driving changes in cyberinfrastructure. In particular, she called out the rapid growth in data from diverse sources, including instruments and sensors and simulation; the increasing complexity of research problems, requiring multidisciplinary teams and multiscale modeling; wider global investment in research, providing more opportunities for collaboration; growing need for technically skilled workforce; and the need for increased societal responsibility and engagement.
NSF has responded to these and other drivers by fielding a diverse array of resources, each of which was spotlighted by one of the panelists:
- Comet, a computing resource focused on the small and medium jobs that represent the “long-tail of science,” at the San Diego Supercomputer Center (SDSC). Comet entered production in May 2015.
- Jetstream, a cloud system with hardware at Indiana University and TACC that is slated to go into production early in 2016.
- Wrangler, a data-intensive system that includes hardware at the Texas Advanced Computing Center (TACC) and Indiana University
- Bridges, a data-centric system slated to go into production early in 2016 at the Pittsburgh Supercomputing Center (PSC).
- Chameleon and CloudLab, testbeds for research on cloud computing.
“I think all of the systems we’re talking about this morning did some interesting and deep analysis of usage patterns” to determine what researchers needed, said Stewart.
For example, SDSC Director Mike Norman said that data from 2012 showed that 99 percent of jobs run on XSEDE-allocated resources used fewer than 2,000 cores and 30 percent used just a single core. Based on that information, SDSC decided to focus Comet on those small to medium jobs, and even to under-allocate the resource so people can get quicker access. They aim to serve 10,000 users per year on Comet, a metric Norman thinks will be easily achieved, in part through embracing Science Gateways.
Jetstream is also aimed at aspects of the long-tail of science, Stewart explained. This cloud system is designed to provide interactive and on-demand computing capabilities via a suite of virtual machines. Users can customize, save, and share VMs—something that Stewart pointed out will make it easier to repeat and reproduce research. And like Comet, Jetstream embraces Science Gateways, working with the iPlant and Galaxy gateways.
A biologist by training, Stewart said that he recently tested the Jetstream interface to see if he could easily “do a little science.”
“It took me about 10 minutes to log in and do something on iPlant and about two hours to do the same thing using Amazon, so the interface really works,” he said.
Both Wrangler and Bridges focus on data needs. Niall Gaffney, director of Data Intensive Computing at TACC, pointed out that traditional high-performance computing systems and ways of working are often mismatched with the needs of data-intensive research. “Databases are not job,” he said. “Scratch is not a storage solution. Hadoop is not always HPC file system-friendly.”
Wrangler is intended to handle big data, lots of small data, structured and unstructured data, and both sequential and random I/O. It also needs to support a large number of applications and interfaces, including Hadoop, Spark, R, GIS, and others.
According to Gaffney, the highly flexible 600 TB flash storage system with bandwidth of 1 TB/sec is one of the most innovative features of Wrangler. “You can connect all 600 TB to one node if that’s what you need,” he said.
As an example of how Wrangler is enabling new data-centric activities, Gaffney said that OrthoMCL, a genomic workflow, would previously not complete on any TACC resource, but now runs in under four hours on Wrangler.
Construction of the data-centric Bridges system will begin in October, according to Nick Nystrom, director of Strategic Applications at PSC. Echoing other panelists, Nystrom agreed that Science Gateways are critical, particularly for communities that are not currently using HPC resources. “Many users don’t want to become programmers,” he said. “Gateways let them avoid a lot of complexity that people associate with traditional supercomputing.”
Bridges will include a pilot project with Temple University, focused on streamlining interoperation and helping people easily move from using campus resources to using nationally available resources such as those provided through XSEDE. “When Temple’s resources are at peak, some jobs can be migrated transparently to Bridges. And conversely, when Bridges is saturated, we can move jobs to Temple,” Nystrom explained.
In addition to these four compute systems available through XSEDE, the panel also highlighted two cloud computing testbeds, Chameleon and CloudLab, which give researchers the opportunity to build and test their own clouds. “There’s still a lot of work to be done in making clouds better and imagining what clouds will look like in the future,” said CloudLab’s Robert Ricci, a research assistant professor at the University of Utah.
The final panelist, Clemson University Jim Bottum, emphasized the need to provide training and assistance so more people from more disciplines can take advantage of all of these diverse computing resources.
“There is a training and education gap between resources and researchers,” he said. “There’s a high barrier to entry without human assistance, and the barriers become higher as we bring in new communities.”
Bottum leads the NSF-supported ACI-REF project, which has begun addressing this gap by enlisting facilitators who can act as “research concierges” for people who are looking for computing resources (or who may not even know what resources are available or how they could impact their research) and by offering training. The goal is to grow the user base, both in terms of the number of people and the number of disciplines using cyberinfrastructure.
After just its first year, ACI-REF’s “concierges” have had 800+ consultations with individual researchers and more than 1,000 people have attended training sessions led by ACI-REF.