Managed process/container

The Cedana daemon is designed to manage the entire lifecycle of a process/container, including checkpoint/restore, in the larger Cedana system.

Managed processes/containers are those that are spawned using cedana run (CLI reference). This command creates a managed job, which can be checkpointed and restored using cedana dump job and cedana restore job subcommands.

By default, jobs are stored in a local DB (in /tmp). You may set the db.path in configuration to change the path to persist it across restarts. If you're authenticated, you may set the db.remote to true in configuration to use a remote DB at your specified Cedana endpoint instead.

Create a new job

To run a new managed job:

cedana run <type> ...

Where <type> can be process, containerd, runc, etc. See features for all plugins that support running managed jobs.

For example, to run a new managed process:

cedana run process --jid <job_id> -- echo "Hello, world!"

The --jid flag is optional, and if not provided, a random job ID will be generated.

Manage an existing job

It's also possible to start managing an existing process/container:

cedana manage <type> ...

Where <type> can be process, containerd, runc, etc. See features for all plugins that support managing existing jobs.

For example, to manage an existing process:

cedana manage process <PID> [args...]

Manage an upcoming job

It's also possible to start managing an upcoming process/container:

cedana manage --upcoming <type> ...

Where <type> can be process, containerd, runc, etc. See features for all plugins that support managing upcoming jobs. The manage command will block and wait for the process/container to become available.

For example, to manage an existing runc container:

cedana manage --upcoming runc <container-id> [args...]

List managed jobs

The subcommand cedana job has many subcommands such as list, kill,m delete, etc. Check the CLI reference for all available subcommands.

cedana job list

JOB             TYPE      PID  STATUS  GPU  CHECKPOINT  SIZE  LOG
used_gould8     process  5336  halted  yes                    /var/log/cedana-output-used_gould8.log
clean_davinci7  runc     3680  halted  no
easy_tu9        process  3892  halted  yes

cedana ps is a shorthand for cedana job list.

Attach I/O

To attach to the I/O of a job, use the --attach flag:

cedana run <type> --attach ...

This will attach the standard input, output, and error streams to the terminal, including the job's exit status. Press Ctrl+C to detach, and the job will continue running in the background.

If you list the jobs, you will see that the job is attachable:

JOB             TYPE      PID  STATUS  GPU  CHECKPOINT  SIZE  LOG
clean_davinci7  runc     3680  halted  no                     [Attachable]

To attach to the job again, use the cedana job attach subcommand:

cedana job attach <job_id>

Or the generic cedana attach if you want to attach using the job's PID:

cedana attach <PID>

If you want to make a job attachable, but not attach to it immediately, you can use the --attachable flag:

cedana run <type> --attachable ...

View job logs

By default, the jobs stdout/stderr are stored in the /var/log/ directory. If you do cedana job list, you will see the path to the log file.

JOB             TYPE      PID  STATUS  GPU  CHECKPOINT  SIZE  LOG
used_gould8     process  5336  halted  yes                    /var/log/cedana-output-used_gould8.log

You can also pass a custom log path. See CLI reference for cedana run subcommand.

If the job has attachable I/O, it will appear as such:

JOB             TYPE      PID  STATUS  GPU  CHECKPOINT  SIZE  LOG
used_gould8     process  5336  halted  yes                    [Attachable]

Checkpoint/restore

Once the daemon has started managing a job, it can be checkpointed and restored using the cedana dump job and cedana restore job subcommands. See managed checkpoint/restore basics for more information.

Checkpoint/restore with GPUs

GPU C/R support is also available for managed jobs. Check out the checkpoint/restore with GPUs guide for more information.