Chapel project meeting: 2025-11-04 (agenda items, meeting link, notes)

[github-actions[bot]]

This discussion is for the weekly Chapel project meeting.

Date/Time: 2025-11-04 at 10am PT
Join the meeting using this Teams link

---

Prior to the meeting, propose any agenda items using comments below. After the meeting, rough notes will be added in comments.

Typically, these meetings consist of:

• News: Quick updates or announcements to keep people in the loop on key efforts or topics
• Demos: Have something cool to show off? Do it in a ~5-minute demo
• Short Topics: Have something that needs a quick discussion, description, or straw poll? Propose a ~5-minute discussion topic
• Longer Topics: If you have a topic needing more time, propose it with an estimated duration and we'll try to fit it in.

If you'd like to propose an item in any of these categories, please reply to this discussion.

In addition to the above sections, we have some standing agenda items, such as:

• Triage: What is the state of testing and are there issues that need more attention?
• User Issues: What new user issues have come up in the past week, and which should we prioritize?
• Q&A: Have something you want to ask, but not on the clock? We'll open the floor to Q&A as the official meeting wraps up.

[jabraham17]

I would like to demo a new flag for vectorization, --vector-library. #27913

[jabraham17]

I'd like to have a brief discussion about #27926 and associated PR #28000 to make sure there are no major objections before proceeding

[DanilaFe]

your two links are the same (both link to the issue)

[bradcray]

I'd like to discuss the addition of an LCM routine to the math module (#18999 and #28005)

[e-kayrakli]

I can give a quick demo on using standard library's sort routine in Arkouda.

[lydia-duncan]

Here's today's user issues report

[bradcray]

Agenda / Notes

News/Announcements

• "Technology Now" podcast episode on Chapel is out as of last Thursday

  • Listen here: https://hpe.lnk.to/Chapelgu

• Brad speaking at SeaGL this Friday

  • https://pretalx.seagl.org/2025/talk/WDYNZF/

• next Tuesday is Veteran's Day -- reschedule this meeting?

Triage Status

• Jade: mid-rotation update
  • Mac testing down due to power outage (Jade)
  • CHAMPS testing broke due to patches requiring changes + ... (Jade)
  • VM testing's been down w/ outages last week, coming back now (Jade/Anna)
  • ongoing sporadic sort cases in Arkouda (Engin)
  • Arkouda IBV correctness testing hitting ZMQ errors (Engin)
    • Engin thinks numeric tests are crashing the server in the previous test
  • Jenkins console authentication is broken (Jade+Anna)
  • comm=ofi jenkins timeouts (Jade)

User Issues

• Lydia: weekly report
  • 0 new user issues
  • closed 1 user issue -- several codes that generate internal errors
    • Daniel changed one case to a user-facing error (de-tupling fields)
  • a few bkmgit PRs open
    • one waiting on changes from him
    • one waiting on design discussions, hopefully in this meeting, driven
      by him

Demos

• Jade: demo '--vector-library' flag (Brad: remember to record?)

  • adding new vectorization flag
  • sample test case doing simple scalar work
  • first compiling in Chapel with --fast and --no-ieee-float
    (to get best vectorization)
  • calls to glibc pow() function, suggesting not getting good vectorization
  • workaround is to pass C flags to specify using a vector library, etc.
    • but not very portable, and required some LLVM/CLANG knowledge
  • so added new '--vector-library' flag
    • e.g., can use to tell compiler we want to use LIBMVEC-X86 (e.g.)
  • can now see vectorized pow calls in binary
  • also using wider vectors now
    • previously were using short vectors and unpacking them to call pow()
  • also works with C back-end
    • illustrates portability challenge -- need different libraries (libmvec)
  • future improvement: Do more to use common names across back-ends
  • Q: Could we imply the flag automatically when using '--fast'?
  • A: Could, but would also typically need to throw '--no-ieee-float'
    • also, "always turning on" doesn't necessarily help, and sometimes hurts
    • Engin: Would like to understand what those regressions are, and how bad
      • (goal: to make user experience simpler / better "out of the box")
  • Q: Are there several such vector libraries, or one per ISA x compiler pair?
  • A: Yes, several.
  • Q: Do you have experience using this flag with others?
  • A: A bit of experience prior to this flag; maybe a tiny bit with it.

• Engin: using the new 'Sort' routine in Arkouda

  • background: have had a Heisenbug related to Arkouda sort crashes on IBV
  • Brad asked question about whether Michael's new sort would help (since
    we may want to use it anyway).
  • Pleasantly, it was essentially a drop-in replacement
  • Seems to resolve bug
  • In terms of performance, seemed blazing-fast compared to previous
  • weak-scaling is so-so, strong scaling looks much better
    • at least to 16 nodes
    • but using a small problem size, which may be part of the issue
  • so nice combination of correctness improvement + perf improvements
  • using ~1TB data set sizes even though 32TB is available
  • running into OOM issues when going much larger
    • sorting uses some temps, but doesn't seem like it should be that much
      Q: What factor of performance between old and new?
      A: Wanted to gather that prior to meeting, but didn't have time; seemed
      several times faster

Discussion Topics

• Jade: CHPL_UNWIND default value

  • Should CHPL_UNWIND be available in the default configuration? #27926
  • Change the default of CHPL_UNWIND to system if possible #28000
  • Currently defaults to none; proposing having it default to something smart
  • Note that information is imprecise/imperfect (e.g., points to proc decl
    point, not where execution is calls are actually made)
  • Any concerns other than portability?
  • Q: Do we risk overwhelming the user with details (like the way Python stack
    traces make me feel)? Could we default to the short form and instruct how
    to re-run enabling it?
  • A: Note that that's somewhat orthogonal to this question since the traces
    can be enabled/disabled at execution-time, but with the default to 'none',
    it's not even an option to get them. But would suggest the reverse,
    defaulting to printing trace with information about how to squash it,
    primarily because you may only get the behavior after a long run, not know
    how to reproduce it, etc.

• Brad: adding Math.lcm()

  • Math module - should we provide lcm as well? Or just point gcd users to BigInteger if they want it? #18999
  • Add first draft of LCM routine #28005
  • Seemingly no concerns about the routine itself
  • Discussion about the need to merge it as @unstable or as part of an
    edition to follow the stability guidelines established after 2.0 release

Post-meeting

Office Hours / Q&A

TODOs / Open Issues

• figure out Veteran's Day conflict
  • we'll skip based on popular demand

[lydia-duncan]

Accidentally tagged the user with the name unstable again 😅