ORNL-Heterogeneous Distributed Computing Research

• Collaboration and “Grid” software
  • Electronic Notebook
  • Cumulvs – remote steering
  • NetSolve – transparent network computing
• Numerical Linear Algebra
  • Scalapack – parallel NLA library
  • ATLAS – automatic NLA optimization
• Heterogeneous Distributed Computing
  • PVM 3.4 – Unix and NT support
  • Harness – Next generation beyond PVM

• Goal: Design a common (open) Notebook Architecture
  • extensible as technology advances
  • interoperable with other notebook viewers
  • customizable for unique inputs of a given project
  • (Web Browser based)
  • Features: Can input anything a web page can hold and includes java applet sketchepad.
  • Can be shared w/ remote collaborators
  • Software is available to set up your own notebook www.epm.ornl.gov/~geist/java/applets/enote/
  • In use by over a 100 groups in research, education, and industry

• Supports Multiple Viewers
• Integrates w/ exiting viz/VR interfaces
  • AVS, vtk, EigenVR, etc.
• Dynamic Attachment and Detachment
• Collaborative Steering of Simulation
• Automatic fault detection
• Checkpoint/Restart of distributed run
• Works with MPI and PVM applications
  • Environment for easy integration of interactive visualization computational steering, and fault tolerance into applications
  • Goal is to accelerate the process of Scientific Discovery
  • Remote steering of a Distributed Application www.epm.ornl.gov/cs/cumulvs.html

• Client-Server design
• Network-enabled solvers
• Seamless access to resources
• Non-hierarchical system
• Load Balancing
• Fault Tolerance
• Heterogeneous Environment Supported
  • Easy-to-use tool to provide efficient and uniform access to a variety of scientific packages on UNIX platforms
  • NetSolve Client
  • NetSolve Agent
  • Network Resources
  • Software Respository
  • Software is available www.cs.utk.edu/netsolve/

• Follow on to LAPACK
• Designed for MPPs – First math lib to do this
• Numerical software that will work on a heterogeneous platform
• Already in use by Cray, IBM, HP-Convex, Fujitsu, NEC, NAG, IMSL
  • Tailor performance & provide support
• HPF interface available
• “Out of Core” implementation
• Still under development,
  • Java version, Sparse direct routines.

• Automatic generation of BLAS for RISC architectures.
• Code generator 6K lines, takes about 1hr to run.
  • Done once for a new architecture.
• Extension of BLAS to Sparse, Parallel and Mixed Precision Operations.
• Extension of ATLAS to higher level operations.
  • SMPs
  • Pentium
  • SGI/Vector
  • DOD DSP
• www.cs.utk.edu/atlas

Vendor: ATLAS

• Tennessee / Oak Ridge Cluster
• Combining Wide-Area Pentium clusters with high-speed networking
• ORNL Myrinet
  • 20x 266Mhz Pentium II
  • Fast Ether Giganet
• UT
  • 8x 300Mhz Pentium II
  • Giganet Myrinet
  • 18x 200Mhz Pro
• ATM
• Mixed Linux/NT

• Computing across mixed Unix and NT clusters
• Communication Context in dynamic environment
  • To allow multiple applications to safely interact
• Persistent messages using tuple space
  • For tool and app discovery and attachment
• Message handlers to allow environment extensions
• User defined tracing to create performance tools
• www.epm.ornl.gov/pvm
• New capabilities for application developers

• Distributed plug-in interface
  • allows user to customize, optimize, and extend the environmentís features to match his applicationís needs.
• Distributed Symmetric Control
  • Ultimate fault tolerance. Mobile applications.
• Multiple distributed virtual machines
  • that can collaborate, merge, or split.
• Dynamic application/tool plug-in
  • allows multiple parallel applications to attach, interact, and detach. Use in our next generation Cumulvs.
  • Pushing the Frontiers of Parallel Virtual Machines
  • www.epm.oml.gov/harness

• Generalized
• Plug-in
• Machine
• Multi-level hot-pluggability allows user to adaptability to a dynamic environment
• Harness
• Dynamically adaptable and extensible VM