Parallel IO

1
Parallel I/O

• Storage Efficient Access Team

2
Application I/O

• Applications have data models appropriate to
  domain
• Multidimensional typed arrays, images composed of
  scan lines, variable length records
• Headers, attributes on data
• I/O system as a whole must
• Provide mapping of application data into storage
  abstractions
• Coordinate access by many processes
• Organize I/O devices into a single space
• And also
• Insulate applications from I/O system changes
• Maintain performance!

3
I/O for Computational Science

• Applications require more software than just a
  parallel file system
• Break up support into multiple layers with
  distinct roles
• Parallel file system maintains logical space,
  provides efficient access to data (e.g. PVFS,
  GPFS, Lustre)
• Middleware layer deals with organizing access by
  many processes(e.g. MPI-IO (ROMIO), UPC-IO)
• High level I/O library maps app. abstractions to
  a structured,portable file format (e.g. HDF5,
  Parallel netCDF)

4
Other Talks

• LBNL - Storage Resource Managers and MPI-IO
• NWU - High-Performance Parallel I/O Libraries
• ORNL - Efficiency of Parallel I/O Software
• PNNL - Active Storage in Parallel File Systems

5
In the Coming Year

• MPI-IO and Extended Attributes (EAs)
• EAs are a relatively new feature of file systems
  that allow additional information to be
  associated with a file or directory
• We will explore using EAs to store information
  useful to MPI-IO and also for enabling access to
  EAs through extensions to the MPI-IO interface

6
In the Coming Year

• Data Models and I/O Interfaces
• A gap still exists between high-level I/O library
  capabilities and APIs and the data models of
  simulation codes
• We will work with application teams to understand
  their data models and how those could be stored
  in existing data formats
• We will develop a bridge API for storing these
  data models in existing HLL formats
• Later we will investigate formats more amenable
  to performance

7
In the Coming Year

• Benchmarking and performance analysis
• Benchmarking of parallel I/O systems is complex,
  and different facets of performance can be
  important depending on the characteristics of the
  applications using the system
• We will continue to evaluate existing tools for
  appropriateness and to fill in gaps where
  necessary to enable appropriate analysis
• Instrumentation of I/O software components may
  play an important role in better understanding
  I/O behavior