The intel_rdt plugin collects information provided by monitoring features of
the Intel Resource Director Technology (Intel(R) RDT). Intel RDT provides the hardware framework to monitor
and control the utilization of shared resources (ex: last level cache, memory bandwidth).
Intel’s Resource Director Technology (RDT) framework consists of:
- Cache Monitoring Technology (CMT)
- Memory Bandwidth Monitoring (MBM)
- Cache Allocation Technology (CAT)
- Code and Data Prioritization (CDP)
As multithreaded and multicore platform architectures emerge, the last level cache and memory bandwidth are key resources to manage for running workloads in single-threaded, multithreaded, or complex virtual machine environments. Intel introduces CMT, MBM, CAT and CDP to manage these workloads across shared resources.
To gather Intel RDT metrics, the intel_rdt plugin uses pqos cli tool which is a
part of Intel(R) RDT Software Package.
Before using this plugin please be sure pqos is properly installed and configured regarding that the plugin
run pqos to work with OS Interface mode. This plugin supports pqos version 4.0.0 and above.
Note: pqos tool needs root privileges to work properly.
Metrics will be constantly reported from the following pqos commands within the given interval:
pqos -r --iface-os --mon-file-type=csv --mon-interval=INTERVAL --mon-core=all:[CORES]\;mbt:[CORES]
where CORES is equal to group of cores provided in config. User can provide many groups.
pqos -r --iface-os --mon-file-type=csv --mon-interval=INTERVAL --mon-pid=all:[PIDS]\;mbt:[PIDS]
where PIDS is group of processes IDs which name are equal to provided process name in a config.
User can provide many process names which lead to create many processes groups.
In both cases INTERVAL is equal to sampling_interval from config.
Because PIDs association within system could change in every moment, Intel RDT plugin provides a functionality to check on every interval if desired processes change their PIDs association. If some change is reported, plugin will restart pqos tool with new arguments. If provided by user process name is not equal to any of available processes, will be omitted and plugin will constantly check for process availability.
Pqos installation process: https://github.com/intel/intel-cmt-cat/blob/master/INSTALL
Enabling OS interface: https://github.com/intel/intel-cmt-cat/wiki, https://github.com/intel/intel-cmt-cat/wiki/resctrl
More about Intel RDT: https://www.intel.com/content/www/us/en/architecture-and-technology/resource-director-technology.html
# Read Intel RDT metrics
[[inputs.intel_rdt]]
## Optionally set sampling interval to Nx100ms.
## This value is propagated to pqos tool. Interval format is defined by pqos itself.
## If not provided or provided 0, will be set to 10 = 10x100ms = 1s.
# sampling_interval = "10"
## Optionally specify the path to pqos executable.
## If not provided, auto discovery will be performed.
# pqos_path = "/usr/local/bin/pqos"
## Optionally specify if IPC and LLC_Misses metrics shouldn't be propagated.
## If not provided, default value is false.
# shortened_metrics = false
## Specify the list of groups of CPU core(s) to be provided as pqos input.
## Mandatory if processes aren't set and forbidden if processes are specified.
## e.g. ["0-3", "4,5,6"] or ["1-3,4"]
# cores = ["0-3"]
## Specify the list of processes for which Metrics will be collected.
## Mandatory if cores aren't set and forbidden if cores are specified.
## e.g. ["qemu", "pmd"]
# processes = ["process"]| Name | Full name | Description |
|---|---|---|
| MBL | Memory Bandwidth on Local NUMA Node | Memory bandwidth utilization by the relevant CPU core/process on the local NUMA memory channel |
| MBR | Memory Bandwidth on Remote NUMA Node | Memory bandwidth utilization by the relevant CPU core/process on the remote NUMA memory channel |
| MBT | Total Memory Bandwidth | Total memory bandwidth utilized by a CPU core/process on local and remote NUMA memory channels |
| LLC | L3 Cache Occupancy | Total Last Level Cache occupancy by a CPU core/process |
| LLC_Misses* | L3 Cache Misses | Total Last Level Cache misses by a CPU core/process |
| IPC* | Instructions Per Cycle | Total instructions per cycle executed by a CPU core/process |
*optional
Pointing to non-existing cores will lead to throwing an error by pqos and the plugin will not work properly. Be sure to check provided core number exists within desired system.
Be aware, reading Intel RDT metrics by pqos cannot be done simultaneously on the same resource. Do not use any other pqos instance that is monitoring the same cores or PIDs within the working system. It is not possible to monitor same cores or PIDs on different groups.
PIDs associated for the given process could be manually checked by pidof command. E.g:
pidof PROCESS
where PROCESS is process name.
> rdt_metric,cores=12\,19,host=r2-compute-20,name=IPC,process=top value=0 1598962030000000000
> rdt_metric,cores=12\,19,host=r2-compute-20,name=LLC_Misses,process=top value=0 1598962030000000000
> rdt_metric,cores=12\,19,host=r2-compute-20,name=LLC,process=top value=0 1598962030000000000
> rdt_metric,cores=12\,19,host=r2-compute-20,name=MBL,process=top value=0 1598962030000000000
> rdt_metric,cores=12\,19,host=r2-compute-20,name=MBR,process=top value=0 1598962030000000000
> rdt_metric,cores=12\,19,host=r2-compute-20,name=MBT,process=top value=0 1598962030000000000