1. RADICAL-Pilot (RP) - Overview

This section provides an overview of RADICAL-Pilot (RP) and related concepts. You will learn what problems RP aims to solve for you. You are introduced to some vocabulary, RP architecture and RP operation.

1.1. What problems does RP solve?

RP supports running applications with many computational tasks on one or more HPC resources, focusing on the following aspects:

  1. Support the efficient concurrent and sequential execution of large numbers of tasks.
  2. Support the execution of heterogeneous tasks. E.g., single-core, OpenMP, MPI, single- and multi-GPU.
  3. Support multiple HPC platforms. E.g., XSEDE HPC resources, DoE and NSF leadership class machines.
  4. Support the execution of tasks on multiple pilots on one or more HPC platforms.

Summarizing, RP is the right tool if, for example, you want to:

  1. perform whatever type of simulation or analysis via a bag of up to tens of thousand of tasks, homogeneous or heterogeneous in size and duration.
  2. write an application/service that requires a pilot-based runtime system capable of executing tasks on one or more HPC platforms, possibly concurrently.

1.2. What is a Task?

Tasks are wrappers around self-contained executables, executed as one or more processes on the operating system of one or more compute nodes of a HPC cluster. As such, tasks are independent, i.e., don’t execute with a shared memory space and tasks are not methods or functions executed by RP on the resources of a HPC platform. Task executables can be any ‘program’ like, for example, Gromacs, Namd or stress but also sleep, date, etc.

1.3. What is a Workload?

Workloads are sets of tasks. RP assumes no priority among the tasks of a workload so workloads are different from workflows. RP makes no assumption about when tasks of a workloads are provided. RP schedule, places and launches the tasks that are available at the moment in which resources are available. For RP it makes no difference if new tasks arrive while other tasks are executing.

1.4. What is a Task (Task)?

In RP, tasks are called Tasks (Task, or ‘task’), indicating that are independent and self-contained tasks of computation. Each Task represents a self-contained, executable part of the application’s workload. A Task is described by the following attributes:

  • executable : the name of the executable to be run on the target machines
  • arguments : a list of argument strings to be passed to the executable
  • environment : a dictionary of environment variable/value pairs to be set before task execution
  • input_staging : a set of staging directives for input data
  • output_staging: a set of staging directives for output data

For more details, see the API documentation

1.5. What is a Pilot?

As an abstraction, a pilot is a placeholder for resources on a given platform and is capable of executing tasks of a workload on those resources. As a system, pilot is a type of middleware software that implements the pilot abstraction.

RP is a pilot system, capable of (1) acquiring resources by submitting jobs to HPC platforms; (2) managing those resources on the user’s (or application’s) behalf; and (3) executing sets and sequences of Tasks on those resources.

Usually, applications written with RP: (1) define one or more pilots; (2) define the HPC platform where each pilot should be submitted; (3) the type and amount of resources that each pilot should acquire on that resource; and (3) the time for which each pilot’s resources should be available (i.e., walltime). Once each pilot is defined, the application can schedule Tasks for execution on it.

Figure 1 shows a high-level representation of RP architecture (yellow boxes) when deployed on two HPC platforms (Resource A and B), executing an application (Application) with 5 pilots (green boxes) and 36 CUs (red circles). Application contains pilot and Task descriptions; RP Client has two components: Pilot Manager and Task Manager. Pilot descriptions are passed to the Pilot Manager and Task descriptions to the Task Manager. The Pilot Manager uses Pilot Launcher to launch 2 of the 5 described pilots. One pilot is submitted to the local Resource Management (RM) system of Resource A, the other pilot to the RM of Resource B. Once instantiated, each pilot becomes available for Task execution. At that point, RP Task Manager sends 2 tasks to Resource A and 5 tasks to Resource B.

RP architecture

Figure 1. High-level view of RP architecture when deployed on a simplified view of two HPC platforms.

1.6. How about data?

Data management is important for executing CUs, both in providing input data, and staging/sharing output data. RP has different means to handle data, and they are specifically covered in sections: in the UserGuide.

1.7. Why do I need a MongoDB to run RP?

RP applications use a MongoDB database to communicate with the pilots they created: upon startup, pilots connect to the MongoDB database and look for CUs to execute. Similarly, pilots push information into the database about, for example, tasks which completed execution. You can run your own MongoDB or use one provided by the RADICAL group. In each case, the MongoDB database needs to be accessible by the login node of the target HPC resource and by the host on which the RP application executes.

Generally, RP applications should not be run on the login node of a HPC machine. RP executes several processes and may require relevant amount of ram, depending on the number of pilots and tasks required by the application. Executing RP on the login node of HPC machines almost certainly violates the policy of fair usage enforced by the managers of those machines. In rare cases, RP has to be executed from a login node but it should not be assumed as the default approach to RP deployment.

More details about MongoDB requirements and deployment can be found in section Installation.

1.8. Why do I need RADICAL-SAGA to run RP?

RP needs to submit one or more jobs to the target HPC machine(s) in order to acquire the resources on which to schedule the application tasks. RP uses SAGA to describe these jobs, independent from the batch system used by each target machine. SAGA is then used to translate this job description into a specific batch job description and to submit it to the machine’s batch system.

1.9. How do I monitor pilots and CUs?

Pilots and tasks progress according to state models. Figure 2 shows the state models of a pilot (left) and of a Task (right). States ending in pending (light blue boxes) indicate that pilots or tasks are queued in one of the RP components. All the other states (blue boxes) indicate that pilots or tasks are managed by an RP component.

Pilot and Task state models.

Figure 2. (left) Pilot state model; (right) Task state model.

When writing an RP application, pilot.state and task.state always report the current state of the entities. Callbacks can be registered for notifications on task and pilot state changes.

Setting the environment variable RADICAL_LOG_LVL=INFO in the shell from which the RP application is executed, turns on logging. Logging provides information about RP’s inner functionality. Pilots running on target resources also create log files, useful for debugging purposes.

1.10. What about logging?

RP supports logging to the terminal and to files. Also, profiles can be written during runtime. You can set the following environment variables in the shell from which the RP application is executed:


The defined verbosity levels are the same as defined by Python’s logging module.

1.11. What about tracing and profiling?

We have tracing capabilities in each RP module, with every time stamp of the trace corresponding to an explicitly defined event of that module or state of the whole RP system. Depending on RP, workload or platform configuration, an event may leave a trace or not at runtime and the order among events may change. States are instead configuration independent, always guaranteed to be traced and always in the given order. As such, states describe the global behavior of RP while events that of each components to which they belong.

Profiling must be explicitly enabled by exporting the following environment variable in the shell from which the RP application is executed:


Further, the argument download=True must be passed to the method session.close(). For more information about profiling see Chapter Profiling.