Monitoring module injects user custom metrics and monitors the process. It supports multiple backends, protocols and data formats.

1. Installation
2. Getting started
3. Advanced features
4. System monitoring and server-side backends installation and configuration

Click here if you don't have aliBuild installed

• Compile Monitoring and its dependencies via aliBuild

aliBuild build Monitoring --defaults o2-dataflow

• Load the environment for Monitoring (in the alice directory)

alienv load Monitoring/latest

Get an instance from MonitoringFactory by passing backend's URI(s) as a parameter (comma separated if more than one). The factory is accessible from o2::monitoring namespace.

#include <MonitoringFactory.h>
using namespace o2::monitoring;
std::unique_ptr<Monitoring> monitoring = MonitoringFactory::Get("backend[-protocol]://host:port[/verbosity][?query]");

See the table below to find URIs for supported backends:

A metric consist of 5 parameters:

• name - metric name
• values - vector of value and value name pairs
• timestamp - time of creation
• verbosity - metric "severity"
• tags - metric metadata represented as map

A metric can be constructed by providing required parameters (value and metric name, value name is set to value):

Metric{10, "name"}

By default metric can be created with zero or one value (in such case value name is set to value). Any additional value may be added using .addValue method, therefore the following two metrics are identical:

Metric{10, "name"}
Metric{"name"}.addValue(10, "value")

1. Metric tags Each metric can be tagged with any number of predefined tags. In order to do so use addTag(tags::Key, tags::Value) or addTag(tags::Key, unsigned short) methods. The latter method allows assigning numeric value to a tag.

Metric{10, "name"}.addTag(tags::Key::Subsystem, tags::Value::QC)

See the example: examples/2-TaggedMetrics.cxx.

1. Global tags Global tags are added to each metric sent eg. hostname tag is added by default by the library.

You can add your own global tag by calling addGlobalTag(std::string_view key, std::string_view value) or addGlobalTag(tags::Key, tags::Value) on Monitoring object.

1. Run number Run number is special case of a global tag, its value can be overwritten at any time, therefore it benefits simplified handling: setRunNumber(uint32_t). Value 0 is unique and means no run number is set.

Pass metric object to send method as l-value reference:

send({10, "name"})
send(Metric{20, "name"}.addTag(tags::Key::CRU, 123))
send(Metric{"throughput"}.addValue(100, "tx").addValue(200, "rx"))

See how it works in the example: examples/1-Basic.cxx.

There are 3 verbosity levels (the same as for backends): Debug, Info, Prod. By default it is set to Verbosity::Info. The default value can be overwritten using: Metric::setDefaultVerbosity(verbosity). To overwrite verbosity on per metric basis use third, optional parameter to metric constructor:

Metric{10, "name", Verbosity::Prod}

Metrics need to match backends verbosity in order to be sent, eg. backend with /info verbosity will accept Info and Prod metrics only.

In order to avoid sending each metric separately, metrics can be temporary stored in the buffer and flushed at the most convenient moment. This feature can be controlled with following two methods:

monitoring->enableBuffering(const std::size_t maxSize)
...
monitoring->flushBuffer();

enableBuffering takes maximum buffer size as its parameter. The buffer gets full all values are flushed automatically.

See how it works in the example: examples/10-Buffering.cxx.

This feature can calculate derived values. To do so, use optional DerivedMetricMode mode parameter of send method:

send(Metric&& metric, [DerivedMetricMode mode])

Two modes are available:

• DerivedMetricMode::RATE - rate between two following values,
• DerivedMetricMode::INCREMENT - sum of all passed values.

The derived value is generated only from the first value of the metric and it is added to the same metric with the value name suffixed with _rate, _increment accordingly.

See how it works in the example: examples/4-RateDerivedMetric.cxx.

This feature provides basic performance status of the process. Note that is runs in separate thread (without mutex).

enableProcessMonitoring([interval in seconds]);

Following metrics are generated every time interval: CPU measurements:

• cpuUsedPercentage - percentage of a core usage (kernel + user mode) over time interval
• involuntaryContextSwitches - involuntary context switches over time interval
• cpuUsedAbsolute - amount of time spent on process execution (in user and kernel mode) over time interval (expressed in microseconds)

Memory measurements: (Linux only)

• memoryUsagePercentage - ratio of the process's virtual memory to memory available on the machine
• virtualMemorySize - virtual memory reserved by process (expressed in kB)
• residentSetSize - resident set size reserved by process (expressed in kB)

Additional metrics are generated at the end of process execution: CPU measurements:

• cpuTimeConsumedByProcess - total amount of time spent on process execution (in user and kernel mode) (expressed in microseconds)
• averageCpuUsedPercentage - average percentage of a core usage over time interval

Memory measurements: (Linux only)

• averageResidentSetSize - average resident set size used by process (expressed in kB)
• averageVirtualMemorySize - average virtual memory used by process (expressed in kB)

[METRIC] <name>,<type> <values> <timestamp> <tags>

The prefix ([METRIC]) can be changed using query component.

Overwrite metric verbosity using regex expression:

Metric::setVerbosityPolicy(Verbosity verbosity, const std::regex& regex)

This guide explains manual installation. For ansible deployment see AliceO2Group/system-configuration gitlab repo.