Applying GPCE Approaches to Distributed and Parallel High-Performance Scientific Computing

1
Applying GPCE Approaches to Distributed and
Parallel High-Performance Scientific Computing

• Nanbor Wangnanbor_at_txcorp.comTech-X Corporation

Workshop on GPCE for QoS Provisioning in
Distributed SystemsPortland, OR, October 23, 2006
Funded by DOE OASCR and OHEP, DoD DARPA and
Navy, and Vanderbilt University
2
Overview of HPC Environments

• Introduction to the problem domains experiences
  from our projects
• The GRID
• OGSA/WSRF framework of services
• Software Installation Management
• HPC Component environment
• Common Component Architecture
• SIDL/Babel language interoperability tool
• DPHPC
• Motivations and approaches
• Remoting components approaches
• Deployment of DPHPC applications
• Computational QoS from ANL

3
HPC Using The Grid

• The GRID
• Making computing resources readily available like
  the power grid to electricity
• Data storage
• CPU cycle
• Globus is the de-facto standard tool
• Example application domains
• HENP data analysis
• Protein folding
• New OGSA (WSRF-based)
• Code generation tools
• Java/C based containers

4
Enhancing the S/W Reusability on the Grid

• Improve job successful rate on the GRID
• Grid Software Installation Management Framework
• Robust/flexible s/w installation
• Based on OGSA services
• Does not address the actual parallel application
  implementations
• Need for modern software development paradigms

5
Common Component Architecture (CCA)

• A component architecture specifically desgined to
  address HPC needs
• Supports scientific programming languages
  (FORTRAN)
• Built-in multi-dimensional array and complex
  supports
• Does not hinder parallel communications
• Core tools
• SIDL/Babel compilers
• Ccaffeine component run-time environment

6
Babel Language Interoperability Tool

• Enable mixing and matching components developed
  in different langauages
• Scientific Interface Definition Language
• Generating Inter-Object Representation and
  specific language mappings
• Also support component implementations
• CCA is defined in SIDL
• Ccaffeine can then load components dynamically

http//www.llnl.gov/casc/components/babel.html
7
Running Parallel CCA Applications
MCMD
SCMD

• Support both SPMD and MPMD scenarios
• Stay out of the way of component parallelism
• Components handle parallel communication

• MPMD can be very complicated to desing
• Also very brittle, non-portalbe, hard to configure

8
Motivations for Mixing Distributed Tech. and
Parallelism

• Provide another high-level abstraction for HPC
  infrastructure
• A new dimension for partitioning application
  compositions
• Motivating example scenarios
• Integrate separately-developed and established
  codes FSP
• Provide a different paradigm for partitioning
  problems multi-physics simulations
• Provide ways to better utilize high-CPU number
  hardware
• Combine computing resources of multiple
  clusters/computing centers
• Enable parallel data streaming between computing
  task and post-processing task
• What we need A Distributed and Parallel
  High-Performance Computing (DPHPC) Environment

9
An Illustration of DPHPC Application

• Still support conventional CCA component managed
  parallelism
• Provide additional framework mediated distributed
  inter-component communication capability

10
Solution Approach

• Goal Mixing distributed and parallel approaches
  to provide a higher composition abstraction
• Existing CCA implementations dont support both
  models
• Demo on integration done before but not part of
  distributions
• Approach
• Connect distributed parallel CCA applications
  using well-accepted tools
• Explore the challenges of developing Distributed
  and Parallel HPC (DPHPC) applications

11
Remoting CCA Components

• Connect distributed parallel computations by
  composing remote-capable proxy components into
  applications
• Remoting component generator
• Babel RMI library
• CORBA mapping
• Other mappingsunder development

12
Examine Deployment Strategies for DPHPC
Applications

• Local-CCA component centric view
• Local applications (implemented benchmarking
  programs for large datasets)
• Employ a distributed builder service for
  registering/requesting distributed ports
• Distributed component centric view
• Two-tier deployment remote components and their
  implementations
• (Beginning to review the latest Data Parallel
  CORBA standard)
• (Brainstormed on implementing a distributed CCA
  runtime using CCM.)
• Grid view
• Making distributed components as grid services

Research on parallel components PACO by Paris
Group in IRISA, France (http//www.irisa.fr/paris/
General/)
13
Computational QoS

• CQoS
• Performance sequential and parallel efficiency
• Accuracy
• Performance monitoring
• Tau (S. Shende, Oregon State)
• Adaptive composition
• Global application context
• Dynamic and adaptive implementation switching

P. Hovland, K. Keahey, L. C. McInnes, B. Norris
(ANL), L. Freitag (Sandia), P. Raghavan (Penn
State)