Benchmarking Working Group

From OpenSFS
Revision as of 08:06, 6 September 2013 by BenEvans (talk | contribs) (File System I/O hero run best practices)
Jump to: navigation, search

Benchmarking Working Group Wiki Pages

Information about the BWG is found at the at website

LUG 2013 OpenSFS BWG Update Presentation

Benchmarking Working Group Tasks

File System I/O hero run best practices

Lead: Ben Evans
Members: Mark Nelson, Ilene Carpenter, Rick Roloff, Nathan Rutman, Liam Forbes

Hero Run Best Practices

I/O Workload Characterization

Lead: Pietro Cicotti
Members: Ilene Carpenter, Rick Mohr, Mike Booth, Ben Evans

Application I/O kernel extraction

Lead: Ilene Carpenter
Members: Jeff Layton, Pietro Ciccotti, Bobbie Lind

File System Monitoring

Lead: Liam Forbes
Members: Alan Wild, Andrew Uselton, Ben Evans, Cheng Shao, Jeff Garlough, Jeff Layton, Mark Nelson, Nic Henke

Metadata Performance Evaluation

Lead: Sorin Fabish
Members: Branislav Radovanovic, Rick Roloff, Cheng Shao, Wang Yibin, Keith Mannthey, Bobbie Lind, Greg Farnum

The task of the BWG Metadata Performance Evaluation Effort (MPEE) group is to:

  1. Build/select tools that will allow evaluation of File System Metadata performance and scalability
  2. The tools will help detect pockets of Metadata low performance in cases when users complain of extreme slowness of MD operations
  3. Benchmark tools will support: POSIX, MPI, and Transactional operations (for CEPH and DAOS)
  4. Address the very high end HPC as well as small and medium installations benchmark needs
  5. Tools applicable to Lustre and: CEPH, GPFS…

Current MPEE proposed list of benchmarks:

  1. mdtest – widely used in HPC
  2. fstest - used by pvfs/OrangeFS community
  3. Postmark and MPI version - old NetApp benchmark
  4. Netmist and MPI version – used by SPECsfs
  5. Synthetic tools – used by LANL, ORNL
  6. MDS-Survey - Intel’s metadata workload simulator.
  7. Any known open source metadata tools used in HPC
  8. Add new Lustre statistics specific to MD operations.

MPEE Usecases

  • mdtest: test file MD operations on MDS: open, create, lookups, readdir; used in academia and as a comparison tool of FS MD.
  • fstest: small I/O’s and small files as well as lookups, targeting both MDS and OSS operations and MD HA for multiple MDS’s.
  • Postmark: old NetApp benchmark – I built an MPI version; it is used to measure MD operations and file size scalability and files per directory scalability.
  • Netmist: used to model any workload from statistics including all MD operations and file operations. Can model Workload objects for I/O performance mixes and combination of I/O and MD. Suitable for initial evaluation of storage as well as for performance troubleshooting.

MPEE Proposed Roadmap

  • Collect benchmark tools candidates from OpenSFS
  • Evaluate all the tools and the workloads that can benchmarked
  • Recommend a small set of MD benchmark tools to cover the majority of MD workloads
  • Collect stats from users of MD benchmarks
  • Build scripts to allow ease of use of the recommended tools
  • Write documentation for troubleshooting MD performance problems using the toolset
  • Create a special website for MD tools

MPEE Asks from OpenSFS

  • Share any open source synthetic benchmarks code
  • Share a list of MD benchmark tools they currently use to allow select the most suitable and used candidates
  • Share MD operations tested to allow build Netmist workload objects
  • Share the MD workloads that create pain points to Lustre FS
  • Share cases of poor MD performance workloads and applications

Return to page.