sub header2

[View PDF]

 

Karsten Schwan, Greg Eisenhauer, Matt Wolf, and Ph.D. students
Georgia Institute of Technology, CERCS Research Center

Applications

  • GTC, GTS, Pixie3D Fusion applications
  • LAMMPS materials modeling code
  • S3D combustion code


Problems

  • Rapid output processing for timely science insights
  • Large I/O output data volume
  • Coupling to science users


Technology Basis

  • ADIOS I/O interface
  • EVPath data streaming middleware as ADIOS transport
  • Location options for locating analytics processing: compute nodes, staging, remote, offline
  • NNTI (Sandia) efficient transport for RDMA


Challenges

  • Limited resources for I/O and analytics
  • High I/O performance with additional online analytics
    • Require online data reduction
    • Require limiting use of disk subsystem
    • Require judicious data movement, analytics placement, and analytics scheduling

 

 

Pixie3D I/O Processing Pipeline

Pixie3D I/O Processing Pipeline

 

Flexible Placement and Execution for In-Situ Analytics


 

Technology Contributions

  • ADIOS/EVPath I/O middleware
    • High performance data movement on IB and UGNI
    • Support diverse in situ analytics placement options
    • Higher-level API: meta-data rich, easy-to-use
  • Flexible Placement
    • Metric-driven optimization, including for end-to-end performance/cost objectives
  • Resource Containers:
    • Resource provisioning for analytics components

Result/Impact

  • Extended ADIOS with new transport to support location-flexible in situ analytics
  • Implemented in situ analytics for GTS, LAMMPS, Pixie3D, S3D
  • Up to 30% end-to-end performance improvement of those applications through flexible placement
  • Utilized DOE-provided NNTI RDMA transport for support of data staging

 

Accelerating Pixie3D I/O Pipeline via Flexible Placement

Schwan.GT.Ana.in.situ.Fusion-fig2

Using 0.78% additional nodes offloading Pixplot and I/O to staging area increases
performance by 33% in comparison to inline placement at the scale of 8192 cores

 

Managing I/O Resources with I/O Containers


 

Applications

  • LAMMPS materials modeling code
  • DOE Sandia applications
  • SmartPointer Scientific Annotation Toolkit

Problems

  • Poor staging resource allocations can cause dataflow bottleneck
  • Complex computational models for analytics execution

Technology Basis

  • ADIOS I/O interface
  • EVPath data streaming for monitoring and control
  • Multilevel management hierarchy
  • Runtime resource management for I/O pipelines
  • Scalable transactions for resilience (with DOE Sandia)

Challenges

  • Limited resources for I/O and analytics
  • Move offline analysis workflows online
    • Must support multiple computational models
    • Provide scalability for non-scalable analysis codes
    • Controlled data movement
    • Provide fault and performance isolation for analysis components and scientific applications

 

Schwan.GT.Ana.in.situ.Fusion-fig3

I/O Containers Overview

 

Managing I/O Resources with I/O Containers


 

Technology

  • I/O Containers
    • Move offline workflows online to operate on data in-transit
    • Runtime resource management to balance resource usage amongst online analysis codes
  • Doubly Distributed Transactions
    • Provide resilience for data movement and control operations in HPC environments

Result/Impact

  • Extended ADIOS and DataTap to use I/O Containers middleware
  • Increased end to end performance for online analysis pipelines
  • Measured performance impact of implementing transactions in HPC environments

 

Improved end to end performance
though I/O Containers management

Schwan.GT.Ana.in.situ.Fusion-fig4

256 simulation nodes; 13 staging nodes

Performance impact of transactions
for control operations

Schwan.GT.Ana.in.situ.Fusion-fig5