cray_cassini

Network performance counter groups for the Cassini NIC

Author:

Hewlett Packard Enterprise Development LP.

Copyright:

Copyright 2021,2023-2025 Hewlett Packard Enterprise Development LP.

Manual section:

5

DESCRIPTION

CrayPat supports experiments that make use of the network performance counters on the Cassini NIC. These counters are accessed through use of the environment variable PAT_RT_PERFCTR and related environment variables. Access to the performance counters on the Cassini NIC is provided by the PAPI cray_cassini component. Execute the command papi_component_avail to determine if this component is supported on your system.

All PAPI utilities such as papi_native_avail can access and list these events. However, when listed via papi_native_avail, the event names contain a cray_cassini::: prefix. This prefix is not required when specifying the event name to CrayPat via the environment variable PAT_RT_PERFCTR or related runtime environment variables.

To list the available events, execute the command papi_native_avail on a compute node and filter for the desired component. For example:

$ srun papi_native_avail -i cray_cassini

For complete lists of the hardware counter events currently supported organized by processor family, execute pat_help and select the counters topic.

By default, counter values for a given Cassini event are aggregated across all Cassini NIC devices attached to the executing process. To acquire the value for a given Cassini event for each NIC device attached to the executing process use the devx modifier when using the PAT_RT_PERFCTR environment variable (this modifier is no longer supported). See the intro_craypat(1) man page for more details.

The Message-Passing Interface (MPI) library supports environment variables that allow access to the Cassini NIC performance counters, although it does not support the PAPI event naming conventions. See the intro_mpi(3) man page for details.

ENVIRONMENT VARIABLES

CRAY_CASSINI_DEBUG

Issues selective debug information about activity in the component during runtime.

Default: 0

CRAY_CASSINI_RAW_COUNTS

Acquires a counter value as-is - the starting value is not subtracted off the acquired value.

Default: Count since start

CRAY_CASSINI_SENSOR_COUNTER_FILE

Specifies the name of the file from which to acquire counter values for the hardware sensor counters.

Default: present_reading

CRAY_CASSINI_SENSOR_TABLE_ENTRIES

Indicates to use the static table entries for the sensort counters in place of the sensor counters collected at runtime.

Default: 0

CRAY_CASSINI_UPDATE_HZ

Frequency at which system call reads are done to acquire the values of selected counter events.

Default: 10 HZ

FILES

/run/cxi/cxi<d>

Contains counter files for the Cassini retry handlers.

/sys/class/cxi_user/cxi<d>/device/sensors/<n>

Contains counter files for the Cassini sensor hardware.

/sys/class/cxi_user/cxi<d>/device/telemetry

Contains counter files for the Cassini NIC hardware.

SEE ALSO

app3(1), intro_craypat(1), pat_build(1), pat_help(1), pat_report(1), pat_run(1)

papi_component_avail(1), papi_native_avail(1)

accpc(5), cray_pm(5), cray_rapl(5), hwpc(5), cray_cassini(5), uncore(5), papi_counters(5)