Joint laboratory focuses on petascale | News | National Center for Supercomputing Applications at the University of Illinois
Joint laboratory focuses on petascale
08.31.10 - Permalink
The University of Illinois at Urbana-Champaign and INRIA, the French national computer science institute, formed the Joint Laboratory for Petascale Computing in June 2009. The Joint Laboratory is based at Illinois and includes researchers from INRIA, Illinois' Center for Extreme-Scale Computation, and NCSA.
"The Joint Laboratory is focusing on the software challenges found in complex high-performance computers. Illinois is simply one of the very few places in the world where all skills are gathered to address the key challenges of sustained petascale computing. We're very excited to be facing those challenges with the University of Illinois," says INRIA's Franck Cappello, who co-directs the Joint Laboratory.
Much of the Joint Laboratory's work is centered on algorithms and software that will run on Blue Waters and other petascale computers. Software developed for Blue Waters is also expected to run efficiently on other large-scale parallel computers.
"Blue Waters will only be a success if the scientific applications and software that run on it can scale to hundreds of thousands of processorswell beyond where they are today," says Marc Snir, who directs the Joint Laboratory with Cappello. Snir is a computer science professor at Illinois and one of the principal investigators on the Blue Waters project.
"There is real urgency to develop the software that will leverage well the powerful Blue Waters platform," he notes. "International collaborations such as the one we set with INRIA help accelerate that work and make sure it is embraced on other systems."
The Joint Laboratory held a workshoplast June in Paris that began looking at:
- Modeling and optimizing numerical libraries, which are at the heart of many scientific applications.
- Fault-tolerance research, which reduces the negative impact when proecessors, disk drives, or memory fail in supercomputers that have tens or hundreds of thousands of those components.
- Novel programming models, which allow scientific applications to be updated or reimagined to take full advantage of extreme-scale supercomputers.
Its second workshop in December at NCSA involved 40 scientists from Illinois and INRIA. Topics discussed at the workshop included the previous as well as Blue Waters hardware and software and hybrid computing. They also identified additional collaborative projects:
- The joint research on combining atomic-level and molecular level simulation will first consist of porting BigDFT in Charm++ and then coupling BigDFT and NAMD. Apart from the computer science aspects, research issues concerning the physics of the coupling will be explored.
- Joint research on resilience and fault tolerance will investigate fundamental properties of HPC applications and derive new fault-tolerance techniques as well as design a new RAS event filtering approach using advanced data-mining techniques.
- Concerning runtime for heterogeneous and hybrid programming, the joint research will concern the management of complex data structures/movements between CPUs and GPUs, process/ thread (re)maping within multi-core cluster nodes, hybrid QCD code and multiple endpoints per MPI process.
- The joint work on MILC (MIMD Lattice Computation) will focus on exploring preconditioners, studying a performance model including memory references for the Fermion force calculation, and investigating multicore implementation strategies.
- Collaborative work on numerical libraries will concern communication-avoiding technique for multicore LU and Dense Eigensystem algorithms for POWER7 and multicore optimizations and communication/computation overlap for 3D-FFT.
- Petascale supercomputers with hundreds of thousands of processes raise serious I/O issues. Initial research will concern I/O scheduling algorithm to stage in-out large data sets from the computing nodes to the file system. Further research will concern an abstraction layer for file access that will control processes access to file system and avoid flash crowd effects (when thousands of processes are opening thousands of files at the same time).
The next workshop will be held in June 2010 in Bordeaux, France. For more information on the Joint Laboratory, including the presentations given at the recent workshop, see: jointlab.ncsa.illinois.edu.