1 / 8

Fault Tolerance BOF

Fault Tolerance BOF. Possible CBHPC paper Co-authors wanted Tammy, Rob, Bruce, Daniel, Nanbor, Sameer, Jim, Doug, David What infrastructure is needed to enable application-level FT in (component) applications? Little experience with anything beyond checkpoint/restart (CR) in general

hagop
Download Presentation

Fault Tolerance BOF

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Fault Tolerance BOF • Possible CBHPC paper • Co-authors wanted • Tammy, Rob, Bruce, Daniel, Nanbor, Sameer, Jim, Doug, David • What infrastructure is needed to enable application-level FT in (component) applications? • Little experience with anything beyond checkpoint/restart (CR) in general • Assume an FT-friendly lower-level environment • Event service for awareness of faults • Ability to request certain behavior from lower-level software • Example: scheduler shouldn’t automatically kill an FT job

  2. Use Cases • MCMD app, 1 node fails • Recovery restarting failed task or ignore failure and go on (self-healing) • MCMD C/R

  3. Checkpoint/Restart Taxonomy • System-level • Eg: BLCR, Cray XT (site option?), but not universally available • Store (complete) memory image to stable storage • Daemon schedules checkpoints • No application (or framework) involvement • Possibly app can request checkpoint • Potential problems: open files, driver state, in-flight messages, etc. • Component i/f to system c/r API • Component support for intelligent reduced checkpointing (MyState interface)

  4. Application-level • Coordinated • Uncoordinated • Causal, Message Logging, etc. • Incremental checkpointing support • Capture component assembly • Checkpoint data component • In-memory (copy or RAID), disk, write-behind, etc., special system services • Quality of Fault Tolerance • What does interface look like? Like RMI • Components to detect faults • Reduced storage (satisfying stability criteria, but not all available data) • Checkpoint-free FT data holders

  5. How to capture/restore state of blackbox components? • Components implement SaveYourself method • Central service invokes SaveYourself on all components that implement it • Serialized data sent to central service for storage • How to restart and restore state? • Not all components will implement SaveYourself • May not have state to store • May be error • Check at start of execution and notify user • RestoreYourself • Specify state (in SIDL file?) and auto-gen serialize/unserialize methods

  6. Another idea • Components register their state data with a central service • Breaks encapsulation and OOness, but not a major violation • Could be higher performance

  7. Recovery • Local restore vs global restore? • Rollback • Need to save execution path at which checkpoint was taken? • Put responsibility on app components? • Extend GoPort abstraction to include save/restore/restart? • Framework tells components to restart • What order? Shouldn’t matter

  8. Paper themes • How components can help with FT • Abstracting FT services into reusable components • Abstracting FT requirements into ports • How components make FT more complicated • Don’t have monolithic view of application (state) • Going beyond CR

More Related