|
| 1 | +# Technical Concepts |
| 2 | + |
| 3 | +The `executorlib` package is designed to up-scale Python functions for High Performance Computing (HPC) by extending the standard Python `Executor` interface. This document explains the underlying technical concepts and the internal architecture of `executorlib`. |
| 4 | + |
| 5 | +## Internal Architecture |
| 6 | + |
| 7 | +The `executorlib` library is structured into four primary modules: |
| 8 | + |
| 9 | +* **`executor`**: Defines the user-facing `Executor` classes (e.g., `SingleNodeExecutor`, `SlurmClusterExecutor`, `SlurmJobExecutor`, `FluxClusterExecutor`, `FluxJobExecutor`). These classes provide the primary interface for users to submit tasks. |
| 10 | +* **`task_scheduler`**: Manages the distribution and scheduling of tasks. It handles task queues, resource allocation, and coordinates with spawners. |
| 11 | +* **`standalone`**: Contains utility functions and classes that do not depend on other internal modules. This includes serialization (using `cloudpickle`), ZMQ-based communication (`SocketInterface`), and input validation. |
| 12 | +* **`backend`**: Contains the code executed by the worker processes to perform the actual function calls. |
| 13 | + |
| 14 | +## Class Hierarchy and Coupling |
| 15 | + |
| 16 | +The following diagram illustrates the relationship between the main classes in `executorlib`. |
| 17 | + |
| 18 | +```{mermaid} |
| 19 | +classDiagram |
| 20 | + class FutureExecutor { |
| 21 | + <<interface>> |
| 22 | + } |
| 23 | + class BaseExecutor { |
| 24 | + -_task_scheduler: TaskSchedulerBase |
| 25 | + +submit(fn, *args, **kwargs) Future |
| 26 | + +shutdown(wait) |
| 27 | + } |
| 28 | + class TaskSchedulerBase { |
| 29 | + -_future_queue: Queue |
| 30 | + -_process: Thread |
| 31 | + +submit(fn, *args, **kwargs) Future |
| 32 | + } |
| 33 | + class BaseSpawner { |
| 34 | + <<interface>> |
| 35 | + +bootup(command_lst) |
| 36 | + +shutdown(wait) |
| 37 | + } |
| 38 | + class SocketInterface { |
| 39 | + +send_dict(input_dict) |
| 40 | + +receive_dict() dict |
| 41 | + } |
| 42 | +
|
| 43 | + FutureExecutor <|-- BaseExecutor |
| 44 | + BaseExecutor o-- TaskSchedulerBase |
| 45 | + TaskSchedulerBase <|-- OneProcessTaskScheduler |
| 46 | + TaskSchedulerBase <|-- BlockAllocationTaskScheduler |
| 47 | + TaskSchedulerBase <|-- DependencyTaskScheduler |
| 48 | + TaskSchedulerBase <|-- FileTaskScheduler |
| 49 | +
|
| 50 | + OneProcessTaskScheduler o-- BaseSpawner |
| 51 | + BaseSpawner <|-- MpiExecSpawner |
| 52 | + BaseSpawner <|-- SrunSpawner |
| 53 | + BaseSpawner <|-- FluxPythonSpawner |
| 54 | +
|
| 55 | + OneProcessTaskScheduler ..> SocketInterface : uses |
| 56 | +``` |
| 57 | + |
| 58 | +## Execution Flow |
| 59 | + |
| 60 | +When a user submits a function to an executor, several steps occur in the background to ensure the task is executed with the requested resources and the result is returned. |
| 61 | + |
| 62 | +```{mermaid} |
| 63 | +sequenceDiagram |
| 64 | + participant User |
| 65 | + participant Executor |
| 66 | + participant TaskScheduler |
| 67 | + participant Spawner |
| 68 | + participant Backend |
| 69 | +
|
| 70 | + User->>Executor: submit(fn, args, resource_dict) |
| 71 | + Executor->>TaskScheduler: submit(fn, args, resource_dict) |
| 72 | + TaskScheduler->>TaskScheduler: Add to _future_queue |
| 73 | + TaskScheduler-->>User: Return Future object |
| 74 | +
|
| 75 | + Note over TaskScheduler, Spawner: Task loop in background thread |
| 76 | +
|
| 77 | + TaskScheduler->>Spawner: bootup(command) |
| 78 | + Spawner->>Backend: Start worker process |
| 79 | + TaskScheduler->>Backend: Send function and arguments (ZMQ/File) |
| 80 | + Backend->>Backend: Execute function |
| 81 | + Backend->>TaskScheduler: Send result (ZMQ/File) |
| 82 | + TaskScheduler->>User: Update Future with result |
| 83 | +``` |
| 84 | + |
| 85 | +## Communication Modes |
| 86 | + |
| 87 | +`executorlib` supports two primary communication modes between the main process and the worker processes: |
| 88 | + |
| 89 | +### Interactive Communication (ZMQ-based) |
| 90 | +Used by `SingleNodeExecutor` and `HPC Job Executor`. It leverages [ZeroMQ (ZMQ)](https://zeromq.org) and [cloudpickle](https://github.com/cloudpipe/cloudpickle) for high-performance, in-memory communication of Python objects. This mode is ideal for low-latency task distribution within an allocation. |
| 91 | + |
| 92 | +### File-based Communication |
| 93 | +Used by the `HPC Cluster Executor`. It uses the filesystem to communicate between the main process and the individual HPC jobs. This mode is necessary when tasks are submitted as independent jobs to a scheduler like SLURM or Flux, where direct network communication between the login node and compute nodes might be restricted. |
| 94 | + |
| 95 | +## Resource Management |
| 96 | + |
| 97 | +One of the key features of `executorlib` is the ability to specify resources on a per-function-call basis using the `resource_dict`. |
| 98 | + |
| 99 | +* **`cores`**: Number of MPI ranks or CPU cores. |
| 100 | +* **`threads_per_core`**: Number of OpenMP threads. |
| 101 | +* **`gpus_per_core`**: Number of GPUs. |
| 102 | +* **`cwd`**: Working directory for the task. |
| 103 | + |
| 104 | +The `TaskScheduler` ensures that these resource requirements are translated into appropriate commands for the `Spawner` (e.g., `mpiexec`, `srun`, or `flux run`). |
![google-labs-jules[bot]](https://avatars.githubusercontent.com/in/842251?v=4&size=40){kind=avatar}
0 commit comments