GitLab

- Allows to instantiate an arbitrary number of sensors which output
incremental values
- Naming and MQTT topic selection are performed automatically
- Useful to stress test the collectagent and verify overall performance

- Added a qosLevel parameter for MQTT communication in the config
- Can either be 0, 1 or 2
- Default is 1, but setting a value of 0 can help improving latency
in case of extremely high message rates

Re-adding the check for errors when opening file descriptors for additional sensors in a group and obtaining IDs. This time properly.

Revert "Check for errors when opening file descriptors for additional sensors in a group and obtaining IDs" as it also has some unintended changes from the supermucng branch

This reverts commit 37e9f37a.

- dcdbpusher can now be configured to automatically publish its
sensors by communicating the MQTT topic -> name mapping to the
collectagent through appropriate messages
- See the README for further details
- As of now, dcdbpusher performs auto-publish (when required to) and
then proceeds with normal operation; this should be changed so that
the pusher terminates gracefully after having published the sensors

Switch to mosquitto_loop instead of mosquitto_loop_start to avoid creating an additional thread.

- Now all sensors across all plugins adopt the naming convention
cpu<CPUid>.<sensorName>, if CPU-related, or just <sensorname> otherwise
- Code to parse cpu sets used both in Perfevent and ProcFS plugins
moved up in ConfiguratorTemplate
- Misc code to format MQTT parts also moved up in ConfiguratorTemplate

- Added a ProcstatPercParser: similar to ProcstatParser, for each CPU-related line it will collect for each column the percentage of the related value with respect to the sum of all columns, instead of raw values
- Can be selected by specifying sensor groups of type "procstat_perc"

- ProcFS plugin now correctly logs MQTT topics for each sensor (debug level)

- Fixed crash when sensor groups cannot be initialized
- Column metrics for each CPU in /proc/stat are now disambiguated from those found in other lines (like softirq) and they are defined with the "col_" prefix

- Now single sensors for distinct metrics can be specified within configuration files for each sensor group
- Major streamlining of code
- Documentation updated to reflect the new config options

- Still work in progress

- The same signal handler is used for both SIGINT and SIGTERM signals
- Previously, the SIGTERM handler would not perform any cleanup,
leading to corruption and issues on the collect agent side

- Most metric name checks in the procstat parser now use plain C string
    matching instead of regular expressions
- Overhead has been reduced to 1.09% for a full vmstat + meminfo +
    procstat instantiation on a CoolMUC3 node (2700+ metrics each sec)
- Overhead for a minimal meminfo instantiation is 0.05% (44 metrics)
- We can safely assume linear scaling of the plugin in function of the
    number of metrics

- Switch from the STD regex library (which apparently is incomplete for GCC < 4.9) to the BOOST one
- Plugin is stable and tested on a CoolMUC3 node
- Current overhead for a full vmstat + meminfo + procstat instantiation is 1.3% on a CoolMUC 3 node with 256 (logical) cores
- More optimization work to come

- Supports sampling of the vmstat, meminfo, and stat files in the proc virtual filesystem
- Basic functionality verified, extensive testing to come in the next days
- Readme and Makefile have been updated as well

Change semantics of MQTT topic for perf events: the mqttpart string in a perfevents group is only used as a place holder for the CPU id and determines the width to be used for the CPU id in the MQTT string.

PDU plugin: moved request attribute from pdu to group; Remove TTL attribute + necessary internal changes