papi_sampler

Man page for the LDMSD papi_sampler plugin

Date:: 30 Sep 2019
Manual section:: 7
Manual group:: LDMS sampler

SYNOPSIS

Within ldmsd_controller or a configuration file: config name=papi_sampler producer=PRODUCER instance=INSTANCE [ component_id=COMP_ID ] [ stream=STREAM ] [ job_expiry=EXPIRY_SEC ]

DESCRIPTION

papi_sampler monitors PAPI events of processes of Slurm jobs.

The job script must define SUBSCRIBER_DATA environment variable as a JSON object that has at least “papi_sampler” attribute as follows:

SUBSCRIBER_DATA='{"papi_sampler":{"file":"/PATH/TO/PAPI.JSON"}}'

where the “file” attribute inside “papi_sampler” points to a JSON-formatted text file containing user-defined schema name and PAPI events of interest, e.g.

{
  "schema": "my_papi",
  "events": [
    "PAPI_TOT_INS",
    "PAPI_L1_DCM"
  ]
}

papi_sampler relies on slurm_notfifier SPANK plugin to notify it about the starting/stopping of jobs on the node over ldmsd_stream. Please consult :ref:`slurm_notifier(7) <slurm_notifier>` for more information on how to deploy and configure it. The value of SUBSCRIBER_DATA from the job script is carried over to papi_sampler when the job started, and an LDMS set will be created according to the PAPI JSON file pointed by the SUBSCRIBER_DATA. In the case of multi-tenant (multiple jobs running on a node), each job has its own set. The set is deleted after job_expiry period after the job exited.

CONFIG OPTIONS

name=papi_sampler: This MUST be papi_sampler (the name of the plugin).
producer=PRODUCER: The name of the data producer (e.g. hostname).
instance=INSTANCE: This is mandatory due to the fact that papi_sampler extends sampler_base and this option is required by sampler_base config. However, the value is ignored and can be anything. The actual name of the papi_sampler instance is PRODUCER/SCHEMA/JOB_ID.
component_id=COMPONENT_ID: An integer identifying the component (default: 0).
stream=STREAM: The name of the stream that slurm_notifier SPANK plugin uses to notify the job events. This attribute is optional with the default being slurm.
job_expiry=EXPIRY_SEC: The number of seconds to retain the set after the job has exited. The default value is 60.

BUGS

No known bugs.

ERRORS

Plugin configuration will fail with ENOTSUP unless the daemon is configured with ‘msg_enable’.

EXAMPLES

Plugin configuration example:

load name=papi_sampler
config name=papi_sampler producer=node0 instance=NA component_id=2 job_expiry=10
start name=papi_sampler interval=1000000 offset=0

Job script example:

#!/bin/bash
export SUBSCRIBER_DATA='{"papi_sampler":{"file":"/tmp/papi.json"}}'
srun bash -c 'for X in {1..60}; do echo $X; sleep 1; done'

PAPI JSON example:

{
  "schema": "my_papi",
  "events": [
    "PAPI_TOT_INS",
    "PAPI_L1_DCM"
  ]
}