Igor (major rewrite)

[bholt]

As a tool that helps you run experiments, the name "Igor" plays on the "Lab Assistant" role.

This branch introduces a major rewrite of the experiments infrastructure. It still relies on experiments.rb for some functionality, but is mostly a full rewrite which resides in igor.rb

New features

• New Ruby DSL syntax for files
• Experiments submitted as batch jobs so they do their own writing to the SQLite database
• "Interactive" mode based on the DSL syntax and Pry's REPL (addressing Issue #2)
• Rebranding as "Igor" ("Lab Assistant", "Interactive Gathering Of Results")
• Object-oriented design

  Currently missing features (in original)

• "Expression" parameters that do calculation with other params

  Targeted features

• Extensible setup (and teardown?) code
  • Grappa (and others) use this to, build, copy over binaries & libraries, etc.
• Better way to combine experiments? (brainstormed about new DSL features to allow multiple sweeps to be combined smoothly)
• Improved handling of params->flags, allow customized mapping from param name to flags
  • for example: the pattern "--flushticks=%{flushticks}" could be automatically filled in for all params
  • might have to be able to define parameter types, and optionally specify both sides for some

[bholt]

A bunch of the early commits may reference a separate "cloister" project. In the end, I failed to get that to automatically marshal arbitrary procs (relied on bad/incomplete 'SerializableProc' gem). So instead, this one works by running a batch script that unmarshals a much simpler Experiment record from disk that has mostly just data, but saves the "parser" proc as a string which is eval'd when the Experiment is unmarshalled.

[bholt]

TODO

• [ ] sbatch stdout/stderr direct to local filesystem and then move back (chkfile?)
• [ ] use local mpirun
• [ ] fix parser (things declared outside...)
• [ ] find other running jobs using slurm FFI
• [ ] better way to tail a job (sattach?)

[bholt]

Mostly done.

Todo:

• [ ] Documentation, lotsa documentation.
• [ ] Be able to find failed jobs.

[bholt]

this is now a new repo: Igor.