executors: fix performance regressions

[arpittkhandelwal]

Description

This Pull Request addresses critical performance regressions in fork_join_executor and parallel_executor by implementing missing traits, optimizing caching strategies, and refactoring the executor template structure.

Key Changes

1. fork_join_executor:
   • Implemented tag_invoke for processing_units_count_t and get_first_core_t. This prevents the executor from falling back to a default (often 1) core count, which was causing algorithms to run sequentially.
   • Added stacksize validation to ensure correct executor instantiation.
     reducing code duplication and maintenance burden.

---

📊 Performance Results

The following results were obtained using the foreach_report_test benchmark with 8 threads in the standard Docker build environment:

Executor	Execution Time (ms)	Status
fork_join_executor	~2.3ms	✅ Fixed (from ~150ms regression)
parallel_executor	~16.8ms	✅ Stable (at parallel baseline)

These results confirm that the sequential fallback regression in fork_join_executor is fully resolved and the parallel_executor optimizations are performing as intended.

---

Closes

This PR supersedes and Closes #6919.

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}

[Copilot]

Pull request overview

This PR targets performance regressions in HPX executors by ensuring fork_join_executor reports correct processing unit traits (avoiding sequential fallbacks) and by optimizing parallel_executor PU-mask handling while reducing template duplication.

Changes:

• Add missing executor trait customizations to fork_join_executor and validate unsupported nostack stacksize.
• Add lazy caching of pu_mask in parallel_policy_executor_base to avoid repeated recomputation in hot paths.
• Unify flat vs hierarchical parallel_policy_executor specializations into a single template with if constexpr and preserve serialization format.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
libs/core/executors/include/hpx/executors/parallel_executor.hpp	Adds PU-mask caching and merges hierarchical/flat executor implementations using if constexpr.
libs/core/executors/include/hpx/executors/fork_join_executor.hpp	Implements missing trait CPOs and rejects thread_stacksize::nostack for correctness.

[hkaiser]

@arpittkhandelwal Would you mind addressing/reacting to the copilot comments, please? I think at least some of them make sense.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

[hkaiser]

While you wrote that you turned mask_ into an atomic, the current state of the PR doesn't relect that.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

[hkaiser]

@arpittkhandelwal thanks for the fixes. However the CIs fail compiling your test.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Comments suppressed due to low confidence (1)

libs/core/executors/include/hpx/executors/fork_join_executor.hpp:1284

• The PR description mentions adding stacksize validation (and the previous docs stated stacksize must not be nostack), but in the current constructors the stacksize value is passed through without validation and the doc note was removed. Either add the promised validation (e.g. reject thread_stacksize::nostack if unsupported) or update the docs/PR description to reflect the actual supported values and behavior (yielding is already conditional on nostack).

        /// \brief Construct a fork_join_executor.
        ///
        /// \param priority  The priority of the worker threads.
        /// \param stacksize The stacksize of the worker threads.
        /// \param sched     The loop schedule of the parallel regions.
        /// \param yield_delay The time after which the executor yields to other
        ///        work if it has not received any new work for execution.
        explicit fork_join_executor(
            threads::thread_priority priority = threads::thread_priority::bound,
            threads::thread_stacksize stacksize =
                threads::thread_stacksize::small_,
            loop_schedule sched = loop_schedule::dynamic,
            std::chrono::nanoseconds yield_delay = std::chrono::microseconds(
                300))
        {

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

[Copilot]

processing_units_count_t customization is introduced with an unconstrained template <typename Parameters>. Other executors constrain this overload to executor_parameters/is_executor_parameters_v, which avoids overly-permissive tag_invoke participation and improves diagnostics. Consider adding the same constraint here to match the rest of the codebase’s patterns for processing_units_count_t.