User Tools

Site Tools


research:projects:siox

SIOX

Scalable I/O For Extreme Performance

The HPC I/O Challenge

During the last decades, increases in processor performance have exceeded those in I/O capacity by a full order of magnitude.

Efforts to bridge this growing gap gave rise to ever more complex I/O architectures, comprising many layers of heterogeneous hard- and software that make performance analysis and localisation of bottlenecks a daunting task at best.

The SIOX Approach

The BMBF-funded project Scalable I/O for Extreme Performance (SIOX) aims to create a system that will aid in and finally automate I/O analysis and optimisation.

To this end, access and performance data will be

  • recorded
    • on all levels
    • in a level-independent format,
  • causally correlated,
  • structured and archived.

SIOX will use the data collected to

  • predict possible performance,
  • evaluate actual performance,
  • locate and analyse bottlenecks and
  • suggest optimized parameters.

The system built is to be deployed on a multitude of parallel file systems running under various operating systems, spanning all major hardware architectures.

Project results will be tested and used with a range of real world climate modelling applications from the DKRZ project portfolio.

Interfaces

All interaction with SIOX is built on the low-level interface siox-ll. Later, the SIOX project will provide more comfortable and well-adapted high-level interfaces for each of the types of logical nodes, such as siox-cache, siox-network or siox-blockstorage. Furthermore, wrappers may be used to easily adapt software already instrumented for other tools to SIOX. Both high-level APIs and wrappers will build upon siox-ll to perform their tasks.

Download

SIOX is an OpenSource project. You can download the source code from the GitHub-Repository or clone the repository using the following command:

git clone https://github.com/JulianKunkel/siox.git

Contact

Project Materials

research/projects/siox.txt · Last modified: 2019-04-26 09:45 by Michael Kuhn

Donate Powered by PHP Valid HTML5 Valid CSS Driven by DokuWiki