geopm_pio_profile(7) – Signals and controls for the ProfileIOGroup

Description

The ProfileIOGroup implements the geopm::IOGroup(3) interface to provide signals and controls from the application based on calls to GEOPM’s prof API geopm_prof(3).

The signals defined herein are useful primarily for end-of-run reporting purposes and for implementing new agents via the geopm::Agent(3) interface.

Requirements

This IOGroup’s signals are only exposed while the GEOPM HPC Runtime is running. The signals are only available to code running within the GEOPM HPC Runtime. The signals cannot be queried via geopmread.

Signals

PROFILE::REGION_HASH

The hash (or ID) of the region of code (see geopm_prof(3)) currently being run by all ranks, otherwise GEOPM_REGION_HASH_UNMARKED. See the enum geopm_region_hash_e defined in geopm_hash.h for more information.

Aggregation: region_hash
Domain: cpu
Format: hex
Unit: hash of region name

PROFILE::REGION_HINT

The region hint (see geopm_prof(3)) associated with the currently running region. For any interval when all ranks are within an MPI function inside of a user defined region, the hint will change from the hint associated with the user defined region to GEOPM_REGION_HINT_NETWORK. If the user defined region was defined with GEOPM_REGION_HINT_NETWORK and there is an interval within the region when all ranks are within an MPI function, GEOPM will not attribute the time spent within the MPI function as MPI time in the report files. It will be instead attributed to the time spent in the region as a whole. See the enum geopm_region_hint_e defined in geopm_hint.h for more information.

Aggregation: region_hint
Domain: cpu
Format: hex
Unit: geopm_region_hint_e

PROFILE::REGION_PROGRESS

Sum of the fractional progress reported by all CPUs over the aggregated domain. When utilizing OpenMP, the per-CPU returned value will be on the interval [0, 1 / num_thread] where num_thread is the number of requested OpenMP threads per rank. When running single-threaded (i.e. not leveraging OpenMP), the per-CPU returned value will either be zero (no active threads), or on the interval [0, 1].

When aggregating this signal to the board domain, the aggregated value will always be on the interval [0, num_rank] where num_rank is equal to the number of ranks (processes) in the run. 0 indicates no progress while 1 (if examining the per-CPU signal) or num_rank (if examining the board aggregated signal) indicates the region is complete.

This signal is not valid outside of a region. Specifically, this signal should be ignored any time the REGION_HASH signal aggregated to the same domain resolves to GEOPM_REGION_HASH_UNMARKED.

Aggregation: sum
Domain: cpu
Format: float
Unit: progress percentage

PROFILE::TIME_HINT_UNKNOWN