cc-metric-collector/collectors/likwidMetric.md

likwid collector

The likwid collector is probably the most complicated collector. The LIKWID library is included as static library with direct access mode. The direct access mode is suitable if the daemon is executed by a root user. The static library does not contain the performance groups, so all information needs to be provided in the configuration.

The likwid configuration consists of two parts, the "eventsets" and "globalmetrics":

• An event set list itself has two parts, the "events" and a set of derivable "metrics". Each of the "events" is a counter:event pair in LIKWID's syntax. The "metrics" are a list of formulas to derive the metric value from the measurements of the "events". Each metric has a name, the formula, a scope and a publish flag. Counter names can be used like variables in the formulas, so PMC0+PMC1 sums the measurements for the both events configured in the counters PMC0 and PMC1. The scope tells the Collector whether it is a metric for each hardware thread (cpu) or each CPU socket (socket). You may specify a unit for the metric with unit. The last one is the publishing flag. It tells the collector whether a metric should be sent to the router.
• The global metrics are metrics which require data from all event set measurements to be derived. The inputs are the metrics in the event sets. Similar to the metrics in the event sets, the global metrics are defined by a name, a formula, a scope and a publish flag. See event set metrics for details. The only difference is that there is no access to the raw event measurements anymore but only to the metrics. So, the idea is to derive a metric in the "eventsets" section and reuse it in the "globalmetrics" part. If you need a metric only for deriving the global metrics, disable forwarding of the event set metrics (publish=false). Be aware that the combination might be misleading because the "behavior" of a metric changes over time and the multiple measurements might count different computing phases. Similar to the metrics in the eventset, you can specify a metric unit with the unit field.

Additional options:

• access_mode : Method to use for hardware performance monitoring (direct access as root user, accessdaemon for the daemon mode)
• accessdaemon_path: Folder with the access daemon likwid-accessD, commonly $LIKWID_INSTALL_LOC/sbin
• force_overwrite: Same as setting LIKWID_FORCE=1. In case counters are already in-use, LIKWID overwrites their configuration to do its measurements
• invalid_to_zero: In some cases, the calculations result in NaN or Inf. With this option, all NaN and Inf values are replaces with 0.0.
• access_mode: Specify LIKWID access mode: direct for direct register access as root user or accessdaemon
• accessdaemon_path: Folder of the accessDaemon likwid-accessD
• liblikwid_path: Location of liblikwid.so

Available metric scopes

Hardware performance counters are scattered all over the system nowadays. A counter coveres a specific part of the system. While there are hardware thread specific counter for CPU cycles, instructions and so on, some others are specific for a whole CPU socket/package. To address that, the collector provides the specification of a 'scope' for each metric.

• cpu : One metric per CPU hardware thread with the tags "type" : "cpu" and "type-id" : "$cpu_id"
• socket : One metric per CPU socket/package with the tags "type" : "socket" and "type-id" : "$socket_id"

Note: You cannot specify socket scope for a metric that is measured at cpu scope, so some kind of expert knowledge or lookup work in the Likwid Wiki is required. Get the scope of each counter from the Architecture pages and as soon as one counter in a metric is socket-specific, the whole metric is socket-specific.

As a guideline:

• All counters FIXCx, PMCy and TMAz have the scope cpu
• All counters names containing BOX have the scope socket
• All PWRx counters have scope socket, except "PWR1" : "RAPL_CORE_ENERGY" has cpu scope
• All DFCx counters have scope socket

Help with the configuration

The configuration for the likwid collector is quite complicated. Most users don't use LIKWID with the event:counter notation but rely on the performance groups defined by the LIKWID team for each architecture. In order to help with the likwid collector configuration, we included a script scripts/likwid_perfgroup_to_cc_config.py that creates the configuration of an eventset from a performance group (using a LIKWID installation in $PATH):

$ likwid-perfctr -i
[...]
short name: ICX
[...]
$ likwid-perfctr -a
[...]
MEM_DP
MEM
FLOPS_SP
CLOCK
[...]
$ scripts/likwid_perfgroup_to_cc_config.py ICX MEM_DP
{
  "events": {
    "FIXC0": "INSTR_RETIRED_ANY",
    "..." : "..."
  },
  "metrics" : [
    {
      "calc": "time",
      "name": "Runtime (RDTSC) [s]",
      "publish": true,
      "unit": "seconds"
      "scope": "cpu"
    },
    {
      "..." : "..."
    }
  ]
}

You can copy this JSON and add it to the eventsets list. If you specify multiple event sets, you can add globally derived metrics in the extra global_metrics section with the metric names as variables.

Mixed usage between daemon and users

LIKWID checks the file /var/run/likwid.lock before performing any interfering operations. Who is allowed to access the counters is determined by the owner of the file. If it does not exist, it is created for the current user. So, if you want to temporarly allow counter access to a user (e.g. in a job):

Before (SLURM prolog, ...)

$ chwon $JOBUSER /var/run/likwid.lock

After (SLURM epilog, ...)

$ chwon $CCUSER /var/run/likwid.lock

Example configuration

  "likwid": {
    "force_overwrite" : false,
    "nan_to_zero" : false,
    "eventsets": [
      {
        "events": {
          "FIXC1": "ACTUAL_CPU_CLOCK",
          "FIXC2": "MAX_CPU_CLOCK",
          "PMC0": "RETIRED_INSTRUCTIONS",
          "PMC1": "CPU_CLOCKS_UNHALTED",
          "PMC2": "RETIRED_SSE_AVX_FLOPS_ALL",
          "PMC3": "MERGE",
          "DFC0": "DRAM_CHANNEL_0",
          "DFC1": "DRAM_CHANNEL_1",
          "DFC2": "DRAM_CHANNEL_2",
          "DFC3": "DRAM_CHANNEL_3"
        },
        "metrics": [
          {
            "name": "ipc",
            "calc": "PMC0/PMC1",
            "type": "cpu",
            "publish": true
          },
          {
            "name": "flops_any",
            "calc": "0.000001*PMC2/time",
            "unit": "MFlops/s",
            "type": "cpu",
            "publish": true
          },
          {
            "name": "clock",
            "calc": "0.000001*(FIXC1/FIXC2)/inverseClock",
            "type": "cpu",
            "unit": "MHz",
            "publish": true
          },
          {
            "name": "mem1",
            "calc": "0.000001*(DFC0+DFC1+DFC2+DFC3)*64.0/time",
            "unit": "Mbyte/s",
            "type": "socket",
            "publish": false
          }
        ]
      },
      {
        "events": {
          "DFC0": "DRAM_CHANNEL_4",
          "DFC1": "DRAM_CHANNEL_5",
          "DFC2": "DRAM_CHANNEL_6",
          "DFC3": "DRAM_CHANNEL_7",
          "PWR0": "RAPL_CORE_ENERGY",
          "PWR1": "RAPL_PKG_ENERGY"
        },
        "metrics": [
          {
            "name": "pwr_core",
            "calc": "PWR0/time",
            "unit": "Watt"
            "type": "socket",
            "publish": true
          },
          {
            "name": "pwr_pkg",
            "calc": "PWR1/time",
            "type": "socket",
            "unit": "Watt"
            "publish": true
          },
          {
            "name": "mem2",
            "calc": "0.000001*(DFC0+DFC1+DFC2+DFC3)*64.0/time",
            "unit": "Mbyte/s",
            "type": "socket",
            "publish": false
          }
        ]
      }
    ],
    "globalmetrics": [
      {
        "name": "mem_bw",
        "calc": "mem1+mem2",
        "type": "socket",
        "unit": "Mbyte/s",
        "publish": true
      }
    ]
  }

How to get the eventsets and metrics from LIKWID

The likwid collector reads hardware performance counters at a cpu and socket level. The configuration looks quite complicated but it is basically copy&paste from LIKWID's performance groups. The collector made multiple iterations and tried to use the performance groups but it lacked flexibility. The current way of configuration provides most flexibility.

The logic is as following: There are multiple eventsets, each consisting of a list of counters+events and a list of metrics. If you compare a common performance group with the example setting above, there is not much difference:

EVENTSET                         ->   "events": {
FIXC1 ACTUAL_CPU_CLOCK           ->     "FIXC1": "ACTUAL_CPU_CLOCK",
FIXC2 MAX_CPU_CLOCK              ->     "FIXC2": "MAX_CPU_CLOCK",
PMC0  RETIRED_INSTRUCTIONS       ->     "PMC0" : "RETIRED_INSTRUCTIONS",
PMC1  CPU_CLOCKS_UNHALTED        ->     "PMC1" : "CPU_CLOCKS_UNHALTED",
PMC2  RETIRED_SSE_AVX_FLOPS_ALL  ->     "PMC2": "RETIRED_SSE_AVX_FLOPS_ALL",
PMC3  MERGE                      ->     "PMC3": "MERGE",
                                 ->   }

The metrics are following the same procedure:

METRICS                          ->   "metrics": [
IPC   PMC0/PMC1                  ->     {
                                 ->       "name" : "IPC",
                                 ->       "calc" : "PMC0/PMC1",
                                 ->       "scope": "cpu",
                                 ->       "publish": true
                                 ->     }
                                 ->   ]

The script scripts/likwid_perfgroup_to_cc_config.py might help you.