ESnet Director Greg Bell Decides Time is Right to Take Next Career Step
After 15 years at Berkeley Lab, including the last six with ESnet, Greg Bell has decided to step down as director of ESnet and the Scientific Networking Division to accept the job of CEO with Broala, a Berkeley company created to support commercial deployment the Bro network monitoring software first developed at the lab. Bell will assume his new job on Monday, Feb. 29.
Bro was first developed by Vern Paxson and has been deployed here since 1996, proving a flexible and highly effective means of protecting the lab from cyber attacks. Paxson is chief scientist for Broala. “I love the mission and people of Berkeley Lab, and the decision to leave wasn’t easy,” Bell said. “But I think we have a chance to succeed in commercializing a truly great technology.” Writing to Computing Sciences staff about Bell’s career change, Associate Lab Director for Computing Sciences Kathy Yelick said “Of course this will be a significant loss to the organization, but this is a great opportunity for Greg and the Lab to spin off a research idea and system into a commercial product, and have a tremendous impact on the broader community.”
In announcing his decision to ESnet staff last week, Bell said “ESnet is in very good shape as an organization. We have talented and dedicated teams, working together cohesively on a vision that’s really about science outcomes - not just networking.” Since he joined ESnet six years ago, the network has deployed a 100 Gigabits-per-second backbone network around the country, and extended its connectivity to Europe with a combined bandwidth of 340 Gbps, a move that Bell said was unprecedented in the research networking community.
ESnet has also become a leader in the growing field of software defined networking, and its Science DMZ architecture for speeding up the flow of large data sets has been adopted by over 100 institutions. “Our ideas have become global best practices for other research and education networks,” said Bell, who was only the fourth person to lead ESnet since the network was established in 1986. “Since we’re just starting to plan for our next major network upgrade, it’s a good moment for someone new to take the reins.” As of Feb. 29, ESnet Chief Technologist Inder Monga will be the acting ESnet director and acting head of the Scientific Networking Division. Yelick will be initiate an international search for the ESnet/division director role.
NERSC's Shifter Highlighted by Cray in HPCwire Article
In an interview with HPCwire about the convergence of HPC and big data, Cray’s Barry Bolding highlighted the company’s Joint Center of Excellence with NERSC, which developed Shifter. Shifter allows HPC users to take advantage of “container-based” computing on systems running Lustre. Container-based services like Docker wrap up a piece of software in a complete filesystem that houses everything it needs to run, including code, runtime, system tools and system libraries. Before Shifter was created, Docker could not be used on Cray systems running Lustre. At SC15, Cray announced it will make Shifter available to all Cray XC customers starting this month, and on the Cray CS400, Cray XE and Cray XK platforms in 2016.
NERSC Supercomputers Aid in Search for New Subatomic Particles
A team of theoretical high-energy physicists in the Fermilab Lattice and MILC Collaborations has published a new, high-precision calculation that could significantly advance the indirect search for physics beyond the Standard Model. The calculation applies to a particularly rare decay of the B meson (a subatomic particle), which is sometimes also called a “penguin decay” process. The team used supercomputing resources at the National Energy Research Scientific Computing Center (NERSC) and other supercomputing facilities across the U.S. to run a number of high-precision lattice QCD calculations and "complete the picture of all things that are possible in the decay of this particular meson," noted Douglas Toussaint, professor of physics at the University of Arizona and the PI who oversees the team’s allocation at NERSC.
This Week's CS Seminars
2015 HPSS User Forum Highlights
Today! Monday, Dec. 7, 12-1 p.m., Bldg. 943, Room 236
Nick Balthaser, NERSC Storage Systems Group
The HPSS User Forum (HUF) is an annual meeting of HPSS developers, admins, project and support staff, and industry representatives involved in shaping the future of HPSS. This year 3 members of the NERSC Storage Systems group attended the 2015 HPSS User Forum held at SciNet in Toronto, ON. In this talk Nick Balthaser summarizes HPSS new features, site and vendor presentations, and group activities from this year’s HUF.
Improving Performance and Manageability of Object-based Storage Systems in HPC
Wednesday, Dec. 9, 11 a.m.-12 p.m., Bldg. 50F, Room 1647 Dong Dai, Texas Tech University Department of Computer Science
Object-based model abstracts data as number of variable size objects with identifiers. This model has been adopted by large number of parallel file systems for referring and accessing data. A high performance and easily managed object-based storage will be the critical part in next-generation (exascale) HPC platforms. In this talk, I will introduce our research in two relevant directions to help build such a storage system: 1) optimized I/O performance through dynamic scheduling; 2) enhanced manageability through extended objects metadata. First, I/O performance especially of burst-write is significantly affected by the short-term stragglers that were generated from interferences among applications and other system components in highly concurrent environment. It is hard to detect or avoid such stragglers as they are irregular and short. In this research, we proposed a new two-choice randomized I/O scheduler that dynamically schedules I/O requests to servers according to their real-time performance. Better load balance and improved overall I/O performance are observed with this new scheduler. Second, HPC systems with object-based storage require advanced data management functionalities that are largely missing in current systems (e.g., result validation, provenance capture). This limits the usability of object storage systems. In this research, we proposed and explored a graph-based rich metadata system, which presents a unified model for enriched metadata; provides scalable metadata read/write performance; and allows users to query their metadata through asynchronous, distributed graph traversal. In addition to improve data management, we further show an example use case of such collected metadata in predicting object access pattern, which in turn improves the I/O performance.