colmet 0.6.10

A utility to monitor the jobs ressources in a HPC environment, espacially OAR

Colmet - Collecting metrics about jobs running in a distributed environnement

Introduction:

Colmet is a monitoring tool to collect metrics about jobs running in a distributed environnement, especially for gathering metrics on clusters and grids. It provides currently several backends :

• Input backends:
  • taskstats: fetch task metrics from the linux kernel
  • rapl: intel processors realtime consumption metrics
  • perfhw: perf_event counters
  • jobproc: get infos from /proc
  • ipmipower: get power metrics from ipmi
  • temperature: get temperatures from /sys/class/thermal
  • infiniband: get infiniband/omnipath network metrics
  • lustre: get lustre FS stats
• Output backends:
  • elasticsearch: store the metrics on elasticsearch indexes
  • hdf5: store the metrics on the filesystem
  • stdout: display the metrics on the terminal

It uses zeromq to transport the metrics across the network.

It is currently bound to the OAR RJMS.

A Grafana sample dashboard is provided for the elasticsearch backend. Here are some snapshots:

Installation:

Requirements

• a Linux kernel that supports

  • Taskstats
  • intel_rapl (for RAPL backend)
  • perf_event (for perfhw backend)
  • ipmi_devintf (for ipmi backend)

• Python Version 2.7 or newer

  • python-zmq 2.2.0 or newer
  • python-tables 3.3.0 or newer
  • python-pyinotify 0.9.3-2 or newer
  • python-requests

• For the Elasticsearch output backend (recommended for sites with > 50 nodes)

  • An Elasticsearch server
  • A Grafana server (for visu)

• For the RAPL input backend:

  • libpowercap, powercap-utils (https://github.com/powercap/powercap)

• For the infiniband backend:

  • perfquery command line tool

• for the ipmipower backend:

  • ipmi-oem command line tool (freeipmi) or other configurable command

Installation

You can install, upgrade, uninstall colmet with these commands::

$ pip install [--user] colmet
$ pip install [--user] --upgrade colmet
$ pip uninstall colmet

Or from git (last development version)::

$ pip install [--user] git+https://github.com/oar-team/colmet.git

Or if you already pulled the sources::

$ pip install [--user] path/to/sources

Usage:

for the nodes :

sudo colmet-node -vvv --zeromq-uri tcp://127.0.0.1:5556

for the collector :

# Simple local HDF5 file collect:
colmet-collector -vvv --zeromq-bind-uri tcp://127.0.0.1:5556 --hdf5-filepath /data/colmet.hdf5 --hdf5-complevel 9

# Collector with an Elasticsearch backend:
  colmet-collector -vvv \
    --zeromq-bind-uri tcp://192.168.0.1:5556 \
    --buffer-size 5000 \
    --sample-period 3 \
    --elastic-host http://192.168.0.2:9200 \
    --elastic-index-prefix colmet_dahu_ 2>>/var/log/colmet_err.log >> /var/log/colmet.log

You will see the number of counters retrieved in the debug log.

For more information, please refer to the help of theses scripts (--help)

Notes about backends

Some input backends may need external libraries that need to be previously compiled and installed:

# For the perfhw backend:
cd colmet/node/backends/lib_perf_hw/ && make && cp lib_perf_hw.so /usr/local/lib/
# For the rapl backend:
cd colmet/node/backends/lib_rapl/ && make && cp lib_rapl.so /usr/local/lib/

Here's acomplete colmet-node start-up process, with perfw, rapl and more backends:

export LIB_PERFHW_PATH=/usr/local/lib/lib_perf_hw.so
export LIB_RAPL_PATH=/applis/site/colmet/lib_rapl.so

colmet-node -vvv --zeromq-uri tcp://192.168.0.1:5556 \
   --cpuset_rootpath /dev/cpuset/oar \
   --enable-infiniband --omnipath \
   --enable-lustre \
   --enable-perfhw --perfhw-list instructions cache_misses page_faults cpu_cycles cache_references \
   --enable-RAPL \
   --enable-jobproc \
   --enable-ipmipower >> /var/log/colmet.log 2>&1

RAPL - Running Average Power Limit (Intel)

RAPL is a feature on recent Intel processors that makes possible to know the power consumption of cpu in realtime.

Usage : start colmet-node with option --enable-RAPL

A file named RAPL_mapping.[timestamp].csv is created in the working directory. It established the correspondence between counter_1, counter_2, etc from collected data and the actual name of the metric as well as the package and zone (core / uncore / dram) of the processor the metric refers to.

If a given counter is not supported by harware the metric name will be "counter_not_supported_by_hardware" and 0 values will appear in the collected data; -1 values in the collected data means there is no counter mapped to the column.

Perfhw

This provides metrics collected using interface perf_event_open.

Usage : start colmet-node with option --enable-perfhw

Optionnaly choose the metrics you want (max 5 metrics) using options --perfhw-list followed by space-separated list of the metrics/

Example : --enable-perfhw --perfhw-list instructions cpu_cycles cache_misses

A file named perfhw_mapping.[timestamp].csv is created in the working directory. It establishes the correspondence between counter_1, counter_2, etc from collected data and the actual name of the metric.

Available metrics (refers to perf_event_open documentation for signification) :

cpu_cycles 
instructions 
cache_references 
cache_misses 
branch_instructions
branch_misses
bus_cycles 
ref_cpu_cycles 
cache_l1d 
cache_ll
cache_dtlb 
cache_itlb 
cache_bpu 
cache_node 
cache_op_read 
cache_op_prefetch 
cache_result_access 
cpu_clock 
task_clock 
page_faults 
context_switches 
cpu_migrations
page_faults_min
page_faults_maj
alignment_faults 
emulation_faults
dummy
bpf_output

Temperature

This backend gets temperatures from /sys/class/thermal/thermal_zone*/temp

Usage : start colmet-node with option --enable-temperature

A file named temperature_mapping.[timestamp].csv is created in the working directory. It establishes the correspondence between counter_1, counter_2, etc from collected data and the actual name of the metric.

Colmet CHANGELOG

Version 0.6.11

Unreleased

Version 0.6.10

• Fixed missing exceptions handling into elasticsearch backend (collector)
• ZMQ: Prefer SNDHWM and RCVHWM to HWM
• Fixed: taskstats data could block other data to be collected when cpuset is empty (node)

Version 0.6.9

• Fix for newer pyzmq versions

Version 0.6.8

• Added nvidia GPU support

Version 0.6.7

• bugfix: glob import missing into procstats

Version 0.6.6

• Added --no-check-certificates option for elastic backend
• Added involved jobs and new metrics into jobprocstats

Version 0.6.4

• Added http auth support for elasticsearch backend

Version 0.6.3

Released on September 4th 2020

• Bugfixes into lustrestats and jobprocstats backend

Version 0.6.2

Released on September 3rd 2020

• Python package fix

Version 0.6.1

Released on September 3rd 2020

• New input backends: lustre, infiniband, temperature, rapl, perfhw, impipower, jobproc
• New ouptut backend: elasticsearch
• Example Grafana dashboard for Elasticsearch backend
• Added "involved_jobs" value for metrics that are global to a node (job 0)
• Bugfix for "dictionnary changed size during iteration"

Version 0.5.4

Released on January 19th 2018

• hdf5 extractor script for OAR RESTFUL API
• Added infiniband backend
• Added lustre backend
• Fixed cpuset_rootpath default always appended

Version 0.5.3

Released on April 29th 2015

• Removed unnecessary lock from the collector to avoid colmet to wait forever
• Removed (async) zmq eventloop and added --sample-period to the collector.
• Fixed some bugs about hdf file

Version 0.5.2

Released on Apr 2nd 2015

• Fixed python syntax error

Version 0.5.1

Released on Apr 2nd 2015

• Fixed error about missing requirements.txt file in the sdist package

Version 0.5.0

Released on Apr 2nd 2015

• Don't run colmet as a daemon anymore
• Maintained compatibility with zmq 3.x/4.x
• Dropped --zeromq-swap (swap was dropped from zmq 3.x)
• Handled zmq name change from HWM to SNDHWM and RCVHWM
• Fixed requirements
• Dropped python 2.6 support

Version 0.4.0

• Saved metrics in new HDF5 file if colmet is reloaded in order to avoid HDF5 data corruption
• Handled HUP signal to reload colmet-collector
• Removed hiwater_rss and hiwater_vm collected metrics.

Version 0.3.1

• New metrics hiwater_rss and hiwater_vm for taskstats
• Worked with pyinotify 0.8
• Added --disable-procstats option to disable procstats backend.

Version 0.3.0

• Divided colmet package into three parts

  • colmet-node : Retrieve data from taskstats and procstats and send to collectors with ZeroMQ
  • colmet-collector : A collector that stores data received by ZeroMQ in a hdf5 file
  • colmet-common : Common colmet part.

• Added some parameters of ZeroMQ backend to prevent a memory overflow

• Simplified the command line interface

• Dropped rrd backend because it is not yet working

• Added --buffer-size option for collector to define the maximum number of counters that colmet should queue in memory before pushing it to output backend

• Handled SIGTERM and SIGINT to terminate colmet properly

Version 0.2.0

• Added options to enable hdf5 compression
• Support for multiple job by cgroup path scanning
• Used Inotify events for job list update
• Don't filter packets if no job_id range was specified, especially with zeromq backend
• Waited the cgroup_path folder creation before scanning the list of jobs
• Added procstat for node monitoring through fictive job with 0 as identifier
• Used absolute time take measure and not delay between measure, to avoid the drift of measure time
• Added workaround when a newly cgroup is created without process in it (monitoring is suspended upto one process is launched)

Version 0.0.1

• Conception