Skip to content

v3.3.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 May 02:14
v3.3.0
This tag was signed with the committer’s verified signature.
rapastranac Andres Pastrana
SSH Key Fingerprint: CgP6JMPsSgxoZqQQtK3ZHNMONanaXcYG8Oh4hyOZAHU
Verified
Learn about vigilant mode.
d57fea1
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v3.3.0

May 21, 2026 · GitHub ↗

Runtime telemetry: process-wide hub with local / TCP / MPI transports, hwloc-backed topology probe, and a small set of read-only runtime accessors. Telemetry ships ON by default; opt out at runtime with gempba::telemetry::disable(). No API breaks.

Added

  • <gempba/telemetry/telemetry_hub.hpp> — process-wide telemetry hub that publishes worker / node frames and routes control messages. Singleton accessed via gempba::telemetry::get(); runtime kill switch via disable() / enable() / is_enabled() (process-local, sticky, must be set symmetrically across MPI ranks)
  • Local in-process and TCP server transports; TCP binds 127.0.0.1:9000 by default (configure_port() to change before the first create_* call)
  • MPI transport on a private communicator (MPI_Comm_dup) so telemetry traffic never collides with application traffic; auto-installed inside mp::create_scheduler so the collective install runs on every rank
  • hwloc-backed topology probe: per-socket physical / logical core counts, total memory, CPU brand, cpu-id list; multi-node topology assembled via MPI_Allgather of the worker_identity POD
  • JSON serializer for telemetry frames; client-pushed interval-control protocol over the TCP socket so dashboards can throttle publish rate live
  • Read-only runtime accessors: gempba::try_get_scheduler (non-throwing variant of get_scheduler), load_balancer::get_thread_pool_size, load_balancer::get_tasks_running_count, scheduler::get_pending_request_count
  • scripts/connect_telemetry.sh — SSH-tunnel helper for inspecting a remote rank's telemetry socket from a local dashboard

Changed

  • Default thread-pool size when no explicit size is set is now 1 (was one-per-core via BS::thread_pool's default). The concrete quasi_horizontal_load_balancer and work_stealing_load_balancer impls construct BS::thread_pool<>{1} explicitly. Users who relied on the implicit default should pass the size explicitly through their scheduler init

Fixed

  • Exported gempbaConfig.cmake now re-discovers hwloc behind GEMPBA_HWLOC, so downstream find_package(gempba) consumers resolve PkgConfig::HWLOC at link time instead of failing with target not found (latent since hwloc became a PRIVATE link dep on a STATIC library)
  • mpi_semi_centralized_scheduler double-counted m_sent_task_count — increments fired on two paths for the same dispatch; consolidated to a single ownership point so the stats visitor and telemetry's record_send see consistent values

Build

  • hwloc is a new runtime dependency, gated by the GEMPBA_HWLOC CMake option (ON for releases, OFF for dev builds). Discovered via pkg-config on all three platforms; .deb and MSYS2 packages declare it as a runtime dep
  • CI installs hwloc on Ubuntu 24.04 / macOS 26 / Windows 2025 (MSYS2) / lint runners
  • build_*.sh scripts forward GEMPBA_HWLOC so packagers can disable hwloc cleanly
  • Windows builds link psapi (process-info probe) and ws2_32 (Winsock for the TCP server)
Assets 4
Loading