G5k-checks: Difference between revisions
Fsaintmarcel (talk | contribs) No edit summary |
|||
Line 337: | Line 337: | ||
Once modifications are tested correct on a maximum of clusters, a new version can be released. | Once modifications are tested correct on a maximum of clusters, a new version can be released. | ||
Rake tasks are provided to ease this process. | Rake tasks are provided to ease this process. You should increase the version number and create a tag | ||
''rake -T | |||
rake package:bump:major # [x.x.x] Increment the major fragment of the version number by 1 | |||
rake package:bump:minor # [x.x.x] Increment the minor fragment of the version number by 1 | |||
rake package:bump:patch # [x.x.x] Increment the patch fragment of the version number by 1 | |||
'' | |||
Then you must push your new version (eg. be careful with no annoted tags and two push commands) | |||
''git push | |||
git push --tags | |||
'' | |||
The | The debian package will be built and published automatically by Gitlab (eg. .gitlab-ci.yaml) | ||
== Environment update == | == Environment update == |
Revision as of 16:50, 9 October 2018
Description
Overview
- g5k-checks is expected to be integrated into the standard environment of the Grid'5000 computational nodes. It checks that a node meets several basic requirements before it declares itself as available to the OAR server.
- This lets the admins enable some checkers which may be very specific to the hardware of a cluster.
Architecture
G5kchecks is based on rspec test suite. Rspec is a little bit roundabout of it first mission: test a program. We use rspec to test all node characteristics. The first step is to retrieve node informatation with ohai. By default ohai provides a large set of characteristics of the machine. Added to this, we have developed some plugins to complete missing information (particularly for the disk, the cpu and the network). The second step is to compare those characteristics with the grid5000 Reference_Repository. To do that, g5kchecks takes each value of the API and compares them with the values given by ohai. If those values don't match, then an error is thrown via the rspec process.
OAR
- The oar-node flavour of OAR installation
/etc/default/oar-node
is started at at boot time. It launches/usr/lib/oar/oarnodecheckrun
, which then runs the executable file/etc/oar/check.d/start_g5kchecks
. The OAR server periodically invokes remotely/usr/bin/oarnodecheckquery
. This command returns with status 1 if/var/lib/oar/check.d/
is not empty, 0 otherwise. So if/etc/oar/check.d/start_g5kchecks
finds something wrong, it simply has to create a log file in that directory. - If
oarnodecheckquery
fails, then the node is not ready to start, and it loops on running those scripts until eitheroarnodecheckquery
returns 0 or a timeout is reached. If the timeout is reached, then it does not attempt to declare the node as "Alive".
This summarizes when g5kchecks is run:
- At service start with
/etc/default/oar-node
- Between (non-deploy) jobs with remote execution of
oarnodecheckrun
andoarnodecheckquery
(In case of deploy jobs, the first type of execution takes place) - Launched by user manually (for now, never happens)
G5kchecks is never run during users jobs.
Checks Overview
The following values are checked by g5k-checks:
# Generated by g5k-checks (g5k-checks -m api) --- genepi-9: architecture: nb_cores: 8 nb_procs: 2 nb_threads: 8 platform_type: x86_64 bios: release_date: 04/30/2008 vendor: Phoenix Technologies LTD version: 1.1a chassis: manufacturer: Bull SAS name: Novascale Universal serial: XAN-SM5P-00187 main_memory: ram_size: 8589934592 network_adapters: bmc: ip: 172.17.16.9 mac: 00:30:48:96:d4:8e management: true eth0: driver: e1000e firmware_version: 2.1-12 interface: Ethernet mac: 00:30:48:7f:3e:ea management: false model: 80003ES2LAN Gigabit Ethernet Controller (Copper) name: enp5s0f0 rate: 0 vendor: Intel eth1: driver: e1000e firmware_version: 2.1-12 interface: Ethernet ip: 172.16.16.9 mac: 00:30:48:7f:3e:eb management: false model: 80003ES2LAN Gigabit Ethernet Controller (Copper) name: enp5s0f1 rate: 1000000000 vendor: Intel ib0: driver: mlx4_core guid: 20:00:55:00:41:80:00:00:00:00:00:00:00:02:c9:03:00:01:5b:15 interface: InfiniBand ip: 172.18.16.9 management: false model: MT26418 [ConnectX VPI PCIe 2.0 5GT/s - IB DDR / 10GigE] rate: 20000000000 vendor: Mellanox Technologies ib1: driver: mlx4_core guid: 20:00:55:00:41:80:00:00:00:00:00:00:00:02:c9:03:00:01:5b:16 interface: InfiniBand management: false model: MT26418 [ConnectX VPI PCIe 2.0 5GT/s - IB DDR / 10GigE] rate: 10000000000 vendor: Mellanox Technologies operating_system: cstate_driver: none cstate_governor: menu cstate_max_id: ht_enabled: false kernel: 3.16.0-4-amd64 name: debian pstate_driver: acpi-cpufreq pstate_governor: performance turboboost_enabled: false version: 8.9 processor: cache_l1d: 32768 cache_l1i: 32768 cache_l2: 6291456 cache_l3: 0 clock_speed: 2500000000 ht_capable: true instruction_set: x86-64 model: Intel Xeon other_description: Intel(R) Xeon(R) CPU E5420 @ 2.50GHz vendor: Intel version: E5420 storage_devices: sda: by_id: "/dev/disk/by-id/wwn-0x50014ee100951626" firmware_version: 20.06C06 model: WDC WD1600YS-01S size: 164696555520 vendor: Western Digital supported_job_types: virtual: ivt
This is an example of output file in API mode (g5k-checks launched with -m api option).
In addition, not all tests are exporting data in this file. The following values are also checked:
- Grid5000 standard environment version
- Grid5000 post-install scripts version
- Usage of sudo-g5k (failed if used, could be destructive to other parts of the system)
- Correct mode of /tmp/
- Fstab partitions mounted and valids
- All partitions have expected size, position, offset, mount options, ...
- Correct KVM driver
Simple usage
Installation
G5kchecks has been tested for jessie, debian stretch is work in progress but should be available soon. On grid5000 debian repository, just add on /etc/apt/sources.list
deb http://apt.grid5000.fr/debian sid main
Install it:
Get sources
git clone https://github.com/grid5000/g5k-checks.git
Run g5k-checks
If you want to check your node just run:
The output should highlight tests in error in red. Also, if some error occured, g5k-checks puts file in /var/lib/g5kchecks/. For instance:
root@adonis-3:~# g5k-checks root@adonis-3:~# ls /var/lib/oar/checklogs/ OAR_Architecture_should_have_the_correct_number_of_thread
You can see the detail of the values checked this way:
root@adonis-3:~# cat /var/lib/oar/checklogs/OAR_Architecture_should_have_the_correct_number_of_thread
Get node description
G5k-checks has a double utility. It can check a node description against our reference API and detect errors. But it can also generate the data to populate this reference API.
If you want a exact node description you can run:
(If launched with -v verbose mode, you can see that almost all tests are failing and it is normal as empty values are checked instead of real ones)
Then g5k-checks put a json and a yaml file in /tmp/
root@adonis-3:~# g5k-checks -m api root@adonis-3:~# ls /tmp/ adonis-3.grenoble.grid5000.fr.json adonis-3.grenoble.grid5000.fr.yaml
Write your own checks/description
G5k-checks internal
G5k-checks is written in ruby on top of the rspec test framework. It gathers informations from ohai program and compare them with grid'5000 reference API data. Rspec is simple to read and write, so you can copy easily other checks and adapt them to your needs.
File tree is:
├── ohai # Ohai plugins, those informations are use by g5k-checks after ├── rspec # Add Rspec formatter (store informations in different way) ├── spec # Checks directory └── utils # some useful class
Play with ohai
Ohai is a small program who retrieve information from different files/other program on the host. It offers an easy to parse output in Json. We can add information to Json just by writing plugins. For instance if we want to add the version of bash in the description, you can create a small file /usr/lib/ruby/vendor_ruby/g5kchecks/ohai/package_version.rb with:
Ohai.plugin(:Packages) do provides "packages" collect_data do packages Mash.new packages[:bash] = `dpkg -l | grep bash | awk '{print $3}'` packages end end
Play with Rspec
Rspec is a framework for testing ruby programs. G5k-checks use Rspec, not to test a ruby program, but to test host. Rspec is simple to read and write. For instance if we want to ensure that bash version is the good one, you can create a file /usr/lib/ruby/vendor_ruby/g5kchecks/spec/packages/packages_spec.rb with :
describe "Packages" do before(:all) do @system = RSpec.configuration.node.ohai_description end it "bash should should have the good version" do puts @system[:packages][:bash].to_yaml bash_version = @system[:packages][:bash].strip bash_version.should eql("4.2+dfsg-0.1"), "#{bash_version}, 4.2+dfsg-0.1, packages, bash" end end
Add checks
Example: I want to check if flag "acpi" is available on the processor:
Add to /usr/lib/ruby/vendor_ruby/g5kchecks/spec/processor/processor_spec.rb:
it "should have apci" do acpi_ohai = @system[:cpu][:'0'][:flags].include?('acpi') acpi_ohai.should_not be_false, "#{acpi_ohai}, is not acpi, processor, acpi" end
Add informations in description
Example: I want to add bogomips of node:
First we should add information in ohai description. To do this we add in the file ohai/cpu.rb after line 80:
if line =~ /^BogoMIPS/ cpu[:Bogo] = line.chomp.split(": ").last.lstrip end
Then we can retrieve information and add it to the description. To do this we add in /usr/lib/ruby/vendor_ruby/g5kchecks/spec/processor/processor_spec.rb:
it "should have BogoMIPS" do bogo_ohai = @system[:cpu][:Bogo] #First value is system, second is from API, thirs is the YAML path in the created '/tmp/' file for -m api mode. #Last argument is false to export value in API mode, true to skip Utils.test(bogo_ohai, nil, 'processor/bogoMIPS', false) do |v_ohai, v_api, error_msg| expect(v_ohai).to eql(v_api), error_msg end end
Now you have the information in /tmp/mynode.mysite.grid5000.fr.yaml:
root@graphene-100:/usr/lib/ruby/vendor_ruby/g5kchecks# g5k-checks -m api root@graphene-100:/usr/lib/ruby/vendor_ruby/g5kchecks# grep -C 3 bogo /tmp/graphene-100.nancy.grid5000.fr.yaml ram_size: 16860348416 processor: clock_speed: 2530000000 bogoMIPS: 5053.74 instruction_set: x86-64 model: Intel Xeon version: X3440
Releasing and testing
Tests and reference-repository update
Before creating a new standard environment, g5k-checks can be tested on target environments using the jenkins test: https://intranet.grid5000.fr/jenkins/job/test_g5kchecks
This test can reserve all or the maximum possible nodes (targets cluster-ALL and cluster-BEST) on each cluster of Grid5000.
It will checkout a (configurable) branch of g5k-checks and test it against a (configurable) branch of the reference-api.
The test will fail if mandatory test fails (i.e. there are entries in /var/lib/oar/checklogs).
Also, the Yaml output of the "-m api" option of g5k-checks will be written to $HOME/g5k-checks-output directory of the ajenkins user on the target site.
Note: it is possible to change the branches of both reference-repository and g5k-checks for the test by configuring the jenkins test:
cd /srv/jenkins-scripts && ruby -Ilib -rg5kcheckstester -e "G5KChecksTester::new.test('$site_cluster', 'dev_feature', 'dev_feature_refrepo')"
For example, this will take the 'dev_feature' branch of g5kcheck and test it against the data present in the 'dev_feature_refrepo' branch of the reference-api.
Updating the reference-repository
Once the tests are finished on the desired clusters, generated Yaml files must be imported manually.
- In the reference repository, go in the generators/run-g5kchecks directory.
- Now get yaml files you want to include. For example:
rsync -va "rennes.adm:/home/ajenkins/g5k-checks-output/paravance*.yaml" ./output/
The output directory hold the temporary files that will be included as input in the reference-repository.
- Run the post-processing script:
ruby ./postprocessing.rb
- Check modified properties of nodes in input:
git diff input/
If values seem correct, generate JSON and commit:
rake reference-api git diff data/ git add data input git commit -m"[SITE] g5k-checks updates"
Release a new version
Once modifications are tested correct on a maximum of clusters, a new version can be released.
Rake tasks are provided to ease this process. You should increase the version number and create a tag
rake -T rake package:bump:major # [x.x.x] Increment the major fragment of the version number by 1 rake package:bump:minor # [x.x.x] Increment the minor fragment of the version number by 1 rake package:bump:patch # [x.x.x] Increment the patch fragment of the version number by 1
Then you must push your new version (eg. be careful with no annoted tags and two push commands)
git push git push --tags
The debian package will be built and published automatically by Gitlab (eg. .gitlab-ci.yaml)
Environment update
The version of g5k-checks included in standard environment is defined in the following file:
virtualbox/steps/data/puppet/modules/env/manifests/std/install_g5kchecks.pp
Once the environment is correct and its version updated, it can be generated with the automated jenkins job: https://intranet.grid5000.fr/jenkins/job/env_generate/
New environment release and reference-api update guidelines
The following procedure summarizes the steps taken to test and deploy a new environment with g5k-checks.
G5k-checks relies on the reference-api to check system data against it. Data from the reference-api must be up-to-date for tests to succeed but most of this data is generated by g5k-checks itself, creating a sort of 'circular dependency'. To avoid dead nodes, g5k-checks data from all nodes should be gathered before pushing a new environment.
- Do a reservation of all nodes of G5K, for example: oarsub -t placeholder=maintenance -l nodes=BEST,walltime=06:00 -q admin -n 'Maintenance' -r '2017-08-31 09:00:00'
The reservation should happen early enough to ensure most (ideally all) of the resources will be available at that time.
- Prepare and release a new debian package of g5k-checks (see #Release a new version)
- Prepare a new standard environment with this new g5k-checks version (see #Environment update)
- Now g5k-checks should be run on all reserved nodes in 'api' mode in order to retrieve the yaml description that will be used to update the reference-api.
This step might be the most tedious one but can be done before the actual deployment. See #Tests and reference-repository update
- Commit and push theses changes to master branch of the reference-repository
- Soon after, push new environment version to all sites using the automated jenkins job: https://intranet.grid5000.fr/jenkins/job/env_push/
The jenkins job does a oar reservation of type 'destructive' that will force the deployment of the new environment.
- If not all nodes were available at the time of the new g5k-checks data retrieval (which is often the case) or during environment update, open a bug # for all sites to let site administrators finish running g5k-checks on remaining nodes.
Run G5k-checks on non-reservable nodes
It is common to update the reference-repository values of nodes whose state are 'Dead' on OAR.
An adaptation of the jenkins g5k-checks test has been made to allow running the same test without doing a OAR reservation.
The only difference is that instead of using OAR to reserve nodes and Kadeploy API to deploy, the nodes are given directly as arguments and kadeploy is called directly from site's frontends.
This scripts must be run on the jenkins machine:
cd /srv/jenkins-scripts
ruby -Ilib -rg5kcheckstester -e "G5KChecksTester::new.from_nodes_list()" grisou-{15,16,18}.nancy.grid5000.fr
Once done, the procedure is the same as described in #Updating the reference-repository.