Getting started¶
This tutorial gets you up and running with a new Balsam Site quickly. Since Balsam is highly platform-agnostic, you can follow along by choosing from any of the available default site setups:
- A local MacOS or Linux system
- Polaris
- Theta-GPU
- Theta-KNL
- Cooley
- Perlmutter
- Summit
- Aurora (coming soon)
Install¶
First create a new virtualenv and install Balsam:
$ /soft/datascience/create_env.sh my-env # Or DIY
$ source my-env/bin/activate
$ pip install --pre balsam
Log In¶
Now that balsam
is installed, you need to log in.
Logging in fetches an access token that is used to identify you
in all future API interactions, until the token expires and you have to log in again.
$ balsam login
# Follow the prompt to authenticate
Once you are logged in, you can create a Balsam Site for job execution, or send jobs to any of your other existing Sites.
Login temporarily restricted
Balsam is currently in a pre-release stage and the web service is hosted on limited resources. Consequently, logins are limited to pre-authorized users. Please contact the ALCF Help Desk to request early access membership to the Balsam user group.
Create a Balsam Site¶
All Balsam workflows are namespaced under Sites: self-contained project spaces belonging to individual users. You can use Balsam to manage Sites on multiple HPC systems from a single shell or Python program. This is one of the key strengths of Balsam: the usage looks exactly the same whether you're running locally or managing jobs across multiple supercomputers.
Let's start by creating a Balsam Site in a folder named ./my-site
:
$ balsam site init ./my-site
# Select the default configuration for Theta-KNL
You will be prompted to select a default Site configuration and to enter a
unique name for your Site. The directory my-site/
will then be created
and preconfigured for the chosen platform. You can list your Sites with the balsam
site ls
command.
$ balsam site ls
ID Name Path Active
21 theta-demo .../projects/datascience/my-site No
In order to actually run anything at the Site, you have to enter the Site
directory and start it with balsam site start
. This command launches
a persistent background agent which uses your access token to sync with the
Balsam service.
$ cd my-site
$ balsam site start
Set up your Apps¶
Every Site has its own collection of Balsam Apps, which define the runnable
applications at that Site. A Balsam App is declared by writing an ApplicationDefinition
class and running the sync()
method from a Python program or interactive session.
The simplest ApplicationDefinition
is just a template for a bash
command, with any workflow variables enclosed in double-curly braces:
from balsam.api import ApplicationDefinition
class Hello(ApplicationDefinition):
site = "theta-demo"
command_template = "echo Hello, {{ say_hello_to }}!"
Hello.sync()
Notice the attribute site = "theta-demo"
which is required to associate the App "Hello"
to the Site "theta-demo"
.
In addition to shell command templates, we can define Apps that invoke a Python run()
function on a compute node:
class VecNorm(ApplicationDefinition):
site = "theta-demo"
def run(self, vec):
return sum(x**2 for x in vec)**0.5
VecNorm.sync()
After running the sync()
methods for these Apps, they are serialized and
shipped into the Balsam cloud service. We can then load and re-use these Apps when
submitting Jobs from other Python programs.
Add Jobs¶
With these App classes in hand, we can now submit some jobs from the Python SDK.
Let's create a Job for both the Hello
and VecNorm
apps: all we need is to
pass a working directory and any necessary parameters for each:
hello = Hello.submit(workdir="demo/hello", say_hello_to="world")
norm = VecNorm.submit(workdir="demo/norm", vec=[3, 4])
Notice how shell command parameters (for Hello
) and Python function parameters
(for VecNorm
) are treated on the same footing. We have now created two Jobs that will eventually run on the Site theta-demo
, once compute resources are available. These Jobs can be seen by running balsam job ls
.
Make it run¶
A key concept in Balsam is that Jobs only specify what to run, and you must create BatchJobs to provision the resources that actually execute your jobs. This way, your workflow definitions are neatly separated from the concern of what allocation they run on. You create a collection of Jobs first, and then many of these Jobs can run inside one (or more) BatchJobs.
Since BatchJobs dynamically acquire Jobs, Balsam execution is fully elastic (just spin up more nodes by adding another BatchJob) and migratable (a Job that ran out of time in one batch allocation will get picked up in the next BatchJob). BatchJobs will automatically run as many Jobs as they can at their Site. You can simply queue up one or several BatchJobs and let them divide and conquer your workload.
BatchJob.objects.create(
site_id=hello_job.site_id,
num_nodes=1,
wall_time_min=10,
job_mode="mpi",
queue="local", # Use the appropriate batch queue, or `local`
project="local", # Use the appropriate allocation, or `local`
)
After running this command, the Site agent will fetch the new BatchJob and
perform the necessary pilot job submission with the local resource manager (e.g.
via qsub
). The hello
and norm
Jobs will run in the workdirs
specified
above, located relative to the Site's data/
directory. You will find the
"Hello world" job output in a job.out
file therein.
For Python run()
applications, the created Jobs
can be handled like
concurrent.futures.Future
instances, where the result()
method delivers the return value (or re-raises
the Exception) from a run()
invocation.
assert norm.result() == 5.0 # (3**2 + 4**2)**0.5
When creating BatchJobs, you can verify that the allocation was actually created
by checking with the local resource manager (e.g. qstat
) or by checking with
Balsam:
$ balsam queue ls
When the BatchJob with job_mode="mpi"
starts, an MPI mode launcher pilot job acquires and runs the jobs. You will find helpful logs in the logs/
directory showing what's going on. Follow the Job statuses with balsam job ls
:
$ balsam job ls
ID Site App Workdir State Tags
267280 thetalogin4:my-site Hello test/2 JOB_FINISHED {}
267279 thetalogin4:my-site Hello test/1 JOB_FINISHED {}
A complete Python example¶
Now let's combine the Python snippets from above to show a self-contained example of Balsam SDK usage (e.g. something you might run from a Jupyter notebook):
from balsam.api import ApplicationDefinition, BatchJob, Job
class VecNorm(ApplicationDefinition):
site = "my-local-site"
def run(self, vec):
return sum(x**2 for x in vec)**0.5
jobs = [
VecNorm.submit(workdir="test/1", vec=[3, 4]),
VecNorm.submit(workdir="test/2", vec=[6, 8]),
]
BatchJob.objects.create(
site_id=jobs[0].site_id,
num_nodes=1,
wall_time_min=10,
job_mode="mpi",
queue="local",
project="local",
)
for job in Job.objects.as_completed(jobs):
print(job.workdir, job.result())
Submitting Jobs from the command line¶
You can check the Apps registered at a given Site, or across all Sites, from the command line:
$ balsam app ls --site=all
ID Name Site
286 Hello laptop
287 VecNorm laptop
To create a Balsam Job from the CLI, you must identify the App on the command line:
$ balsam job create --site=laptop --app Hello --workdir demo/hello2 --param say_hello_to="world2"
There are many additional CLI options in job creation, which can be summarized with balsam job create --help
. You will usually create jobs using Python, but the CLI remains useful for monitoring workflow status:
$ balsam job ls
ID Site App Workdir State Tags
501649 laptop Hello test/1 PREPROCESSED {}
501650 laptop Hello test/2 STAGED_IN {}
BatchJobs can be submitted from the CLI, with parameters that mimic a standard scheduler interface:
# Substitute -q QUEUE and -A ALLOCATION for your project:
$ balsam queue submit -q debug-cache-quad -A datascience -n 1 -t 10 -j mpi