Traceable performance evaluation

From Grid5000
Jump to: navigation, search
Note.png Note

This page is actively maintained by the Grid'5000 team. If you encounter problems, please report them (see the Support page). Additionally, as it is a wiki page, you are free to make minor corrections yourself if needed. If you would like to suggest a more fundamental change, please contact the Grid'5000 team.

This page presents good practices when performing performance evaluations (benchmarks, ...) using Grid'5000 nodes. It is not about benchmarking Grid'5000 itself.

About the Grid'5000 nodes' operating system

Grid'5000 nodes can be used with a choice of different system environments: either with the default (standard) environment which is pre-provisioned on nodes or with a variety of environments deployed on demand by the user. We compare the several options.

Using a node running the default (standard) environment

In this case, access to the system of the node does not require a deployment (kadeploy), so access is quicker. The system however comes with a lot of functionalities that may impact benchmarks for instance. It may be relevant to uninstall/deactivate them before running a performance evaluation.

The list of services running on the system can be shown using systemctl. Then the system can be modified using sudo-g5k to get the root privileges.

Services can be deactivated (with sudo-g5k systemctl stop service). More generally, software packages can be uninstalled (e.g. with sudo-g5k apt-get remove package → any associated services should be stopped, but mind checking).

  • prometheus-node-exporter.service (Prometheus exporter for machine metrics) is typically known to cause power consumption spikes every 15s.
  • dcgm-exporter.service (NVIDIA DCGM prometheus exporter service), which is active (running) on machines equipped with GPUs, can also be unactivated.

Also, the Grid'5000 default environment includes some common system tweaks that may have biases on benchmarks. The CPU parameters page provides relevant information.

Using a deployed environment

Using kadeploy is very relevant in the context of performance evaluation, as it allows a user to operate the nodes with the most minimalistic system that is sufficient for the use case. For instance, one can deploy the -min variant of one of the Grid'5000 supported environments, or a customized environment possibly built on top of a minimalistic environment and featuring only software required for running a benchmark.

Note.png Note

By design, the -min variant of the Grid'5000 supported environments do not change the default behavior of the Linux distribution it is based on.

The recipes from which Grid'5000 environment are built are available in https://gitlab.inria.fr/grid5000/environments-recipes. See Environment creation for more information about how environments are built.

About the traceability of the infrastructure

Special effort is made in Grid'5000 to keep track of infrastructure changes so that any user can get information about for instance what side effects could have impacted an experiment.

The following sources of information are available:

News
The most notable changes in the infrastructure are the subject of news published on the News page as well as on the Grid'5000 twitter and sent to the users' mailing list (users@list.grid5000.fr). It is strongly recommended to follow the news.
Environment recipes
Changes in the way the operating system of nodes is built are tracked in the environment recipes git repository: https://gitlab.inria.fr/grid5000/environments-recipes.
  • It's a good idea to keep track of the version of the system running for an experiment, by storing the information provided in /etc/grid5000/release for the standard environment or other Grid'5000 supported environment. That file contains the name of the environment suffixed with its version (as shown by the kaenv3 command on frontends), as well as the SHA ID of the git commit from which the environment was built. In the git repository, tags mark the commits used for building each version of the supported environments as well. Thus, older environments (used for a previous experiment) can be redeployed on nodes with kadeploy3 in order to help reproduce results.
  • For Grid'5000 supported environments, it is also worth storing the environment postinstall (g5k-postinstall command and version) that was used, which is provided in the /etc/grid5000/postinstall file. The git repository of g5k-postinstall is https://gitlab.inria.fr/grid5000/g5k-postinstall.
Reference repository & Reference API
Detailed information and related changes in the components of the infrastructure are recorded in the reference repository which is served by the Grid'5000 Reference API (REST API). Updates of firmware versions for instance are tracked there. It is a good idea to keep track of the state of the platform at the time of the experiment as explained in API_tutorial#Platform_state_and_reproducibility.

Finally, a new tool currently in alpha version allows for bundling infrastructure metadata information related to an experiment. See Grid5000 Metadata Bundler for more details.

About the Grid'5000 node monitoring service

Grid'5000 provides the kwollect service that can collect information about nodes during a job: energy, network traffic, and more. See the kwollect page or the Energy consumption monitoring tutorial.

Note.png Note

Mind that if you deactivate the Prometheus exporter services on a node (as discussed above), data will stop being gathered for that node in kwollect.