Random module deprecations and stabilization improvements

[jeremiah-corrado]

Stabilization changes to the Random module:

• Deprecates RandomStreamInterface
• Deprecates the RNG, defaultRNG and RandomStream symbols used to select between PCG and NPB rngs
• Deprecates standalone methods that accept an RNG argument in favor of those that don't
• Adds a new record, randomStream which wraps the PCGRandomStream implementation in a slightly simplified interface:
  • a randomStream is now type-specific (can only generate values of one type)
  • a randomStream is not parallel safe
  • fillRandom is now called fill
• Deprecates PCGRandom submodule in favor of usingrandomStream
• Moves the NPBRandom submodule to a standalone package module
• Deprecates createRandomStream in favor of new randomStream
• Deprecates SeedGenerator (and the RandomSupport submodule in general)

The Random module is no longer unstable as a whole. The following symbols are unstable pending further design discussions (summarized here):

• permutation and randomStream.permutation
• randomStream.choice
• randomStream.getNth and randomStream.skipToNth
• randomStream.iterate
• The PCGRandomLib submodule
• methods that internally generate a random seed value

Notes to reviewer

• A majority of the changes in this PR are adjustments to tests to use randomStream instead of PCGRandomStream or NPBRandomStream
  • in most cases randomStream is used (which sometimes involved adjusting .good files when moving from the NPB algorithm)
  • in a few other cases the NPB algorithm is used via the NPBRand package module. This was only done for some performance tests that rely on NPB's relative speed, or in cases where adjusting the good file would not have been trivial.
• Note that many of the tests that use Random are not testing the module itself, but are using it to create random test data or for other algorithms that require randomness (i.e., some sorting algorithms).
  • Only the new tests under deprecated/ or under library/standard/Random/ are testing the module itself and thus should probably be reviewed more closely

Testing:

• local paratest (gnu)
• local paratest (llvm)
• gasnet paratest
• inspected docs

[mppf]

I haven't looked at the test changes yet, but wanted to get my review comments about the module changes to you before I get to the test changes.

[mppf]

Thanks! I'm worried that the tests moved to test/deprecated/random were moved with too big of a brush. Many of these are useful tests of determinism or statistical properties and I think we want to keep those tests. But, I see that many of these tests might use a deprecated feature. If you don't feel you have time to update them before merging this PR, that is OK, but I'd like to see follow-up to move them back to test/library & potentially have the .good files include the deprecation warning until we update them individually to avoid deprecated features.

[jeremiah-corrado]

Thanks for the review @mppf!

Based on your comments above, I'll have two follow up PRs:

• one to update the tests in library/standard/Random/ that have deprecation warnings in their good files. I'll make those use the undeprecated symbols and add new corresponding tests back to /deprecated/random/ as needed.
• another to create a future for the compiler bug I worked around in 6076ab9 (I'll also make an issue for that)

[bradcray]

Hi @jeremiah-corrado and @mppf — I didn’t catch this until now, but I don’t think the HPCC benchmarks should use NPBRandom, but Random instead. The only tests that I’m aware of that care what the random number generator is are the NAS Parallel Benchmarks (where the NPB comes from). All others should focus on using the default/standard random number generator instead to be less "weird" / rely on unstable packages and fetures. For example, stream just wants some data in the vectors, and doesn’t really care what that data is.

Can you take a cleanup item (post-release at this point) to update such cases to simply use Random? If you run into surprises or cases that seem to contradict what I'm saying, please let me know.

[jeremiah-corrado]

Yeah, I can create a follow-on PR next sprint that makes those changes.