Nanos6 model

The Nanos6 runtime library implements the OmpSs-2 tasking model, which schedules the execution of tasks with dependencies. For more information see the OmpSs-2 website and the Nanos6 repository.

The library is instrumented to track the execution of tasks and also the execution path inside the runtime library to identify what is happening. This information is typically used by both the users and the developers of the Nanos6 library to analyze problems and unwanted behaviors. Towards this goal, five different Paraver views are generated, which are explained below.

The state of each task is modelled in a simple finite state machine, which identifies the main state changes of the task. The task is set to the Running state when it begins executing the body of the task, consisting of user defined code. The states can be observed in the following diagram:

A thread can only execute one task at a time, but multiple tasks can be nested and only the topmost one will actually run. When a view presents the running task of a thread, that one is always the top of the stack.

Task ID view

The task ID view represents the numeric ID of the Nanos6 task that is currently running on each thread. The ID is a monotonically increasing identifier assigned on task creation. Lower IDs correspond to tasks created at an earlier point than higher IDs.

Task type view

Every task in Nanos6 contains a task type, which roughly corresponds to the actual location in the code a task was declared. For example if a function is declared as a Nanos6 task, and it is called multiple times in a program, every created task will have a different ID, but the same type.

In the view, each type is shown with a label declared in the source with the label attribute of the running task. If no label was specified, one is automatically generated for each type.

Note that in this view, the numeric event value is a hash function of the type label, so two distinct types (tasks declared in different parts of the code) with the same label will share the event value and have the same color.

MPI rank view

This view shows the numeric MPI rank of the process running the current task. It is only shown when the task is in the running state. This view is specially useful to identify task in a distributed workload, which spans several nodes.

As the zero value in Paraver gets hidden, we use the rank+1 value instead. Therefore the rank numeric value go from 1 to the number of ranks (inclusive).

Thread type view

This view shows the type of each thread:

Main: the first thread executed before the main() function even begins.
Leader: helps with the execution of the main().
Worker: the most common threads, these are in charge of queueing and running tasks.
External: used for external threads that attach to Nanos6 (currently there are not in use).

Idle view

The idle view shows when CPUs become Idle. This state is displayed when no thread is Running in the CPU or when a worker is marked as Stalled (not making progress) such as when busy waiting but still Running in the CPU. The CPUs can also be marked as Absorbing noise, when they are running a worker in the sponge mode, which leaves the CPU free to absorb noise from system processes and interruptions to avoid disruption in the other CPUs.

In particular, a worker requesting a new task will become Stalled after entering the delegation lock and performing a complete iteration without receiving work.

Subsystem view

The subsystem view attempts to provide a general overview of what Nanos6 is doing at any point in time. This view is more complex to understand than the others but is generally the most useful to understand what is happening and debug problems related with Nanos6 itself.

The view shows the state of the runtime for each thread (and for each CPU, the state of the running thread in that CPU).

The state is computed by the following method: the runtime code is completely divided into sections of code (machine instructions) $S_1, S_2, \ldots, S_N$, which are instrumented (an event is emitted when entering and exiting each section), and one common section of code which is shared across the subsystems, `U`, of no interest. We also assume any other code not belonging to the runtime belongs to the `U` section.

!!! remark

Every instruction of the runtime belongs to *exactly one section*.

To determine the state of a thread, we look into the stack to see what is the top-most instrumented section.

At any given point in time, a thread may be executing code with a stack that spawns multiple sections, for example $[ S_1, U, S_2, S_3, U ]$ (the last is current stack frame). The subsystem view selects the last subsystem section from the stack ignoring the common section `U`, and presents that section as the current state of the execution, in this case the section `S_3`.

Additionally, the runtime sections `S_i` are grouped together in subsystems, which form a closely related group of functions. When there is no instrumented section in the thread stack, the state is set to No subsystem. The complete list of subsystems and sections is shown below.

Task subsystem: Controls the life cycle of tasks
- Running body: Executing the body of the task (user defined code).
- Running task for: Running the body of the task in a task for (one of the collaborators).
- Spawning function: Spawning a function as task that will be submitted for later execution.
- Creating: Creating a new task via nanos6_create_task
- Submitting: Submitting a recently created task via nanos6_submit_task
Scheduler subsystem: Queueing and dequeueing ready tasks.
- Serving tasks: Inside the scheduler lock, serving tasks to other workers.
- Adding ready tasks: Adding tasks to the scheduler queues, but outside of the scheduler lock.
- Processing ready tasks: Moving tasks from the lock-free add queues to the scheduler queues.
Worker subsystem: Actions that relate to worker threads, which continuously try to execute new tasks.
- Looking for work: Actively requesting tasks from the scheduler, registered but not holding the lock.
- Handling task: Processing a recently assigned task.
- Switching to another thread: Switching to another thread via switchTo().
- Migrating CPU: Changing the CPU of the current worker.
- Suspending thread: The current thread is in the suspend() function.
- Resuming another thread: Inside the resume() function, wakening another thread.
Memory: Manages the allocation and deallocation of memory.
- Allocating: Allocating new memory. This can take very long with jemalloc in the first calls from each thread.
- Freeing: Deallocating memory.
Dependency subsystem: Manages the registration of task dependencies.
- Registering: Registering dependencies of a task
- Unregistering: Releasing dependencies of a task because it has ended
Blocking subsystem: Code that stops the thread execution.
- Taskwait: Task is blocked due to a taskwait clause
- Blocking current task: Task is blocked through the Nanos6 blocking API
- Unblocking remote task: Unblocking a different task using the Nanos6 blocking API
- Wait for deadline: Blocking a deadline task, which will be re-enqueued when a certain amount of time has passed

!!! remark

Notice that tasks in the running state will display the "Task: Running
body" subsystem only when the task has not called any other instrumented
subsystem in Nanos6. Tasks will continue to be in the running state until
paused or finished.

Breakdown view (experimental)

The breakdown view shows the number of CPUs in a given state, stacking similar states together. Its a mix of the task type, subsystem and idle views.

!!! Important

This view is only generated with the `-b` option of ovniemu. This
feature is experimental, new changes may break compatibility.

Here is an example of the Heat mini-app.

The Idle state shows when a CPU has no threads running or when the worker is Stalled (is not making progress). Below the Idle state the tasks are shown by their label, in this case only block computation is running. Above the Idle state, the runtime subsystems are shown.

This view gives a quick overview of the usage of the CPUs over time and which tasks are taking more time. It also shows which sections are unable to use all the CPUs, showing potential problems with the parallelization. In this example not all the CPUs can be filled with block computation tasks.

Limitations

The task for clause is partially supported, as currently the emulator uses a simplified tasking model where a task can only be executed by one thread, and only once. However, the Nanos6 runtime can run the body of a single task multiple times by varying the arguments with the task for clause, which breaks the emulation model.

The task for is currently only shown in the subsystem view, but it doesn't appear as running any task in the other views.

9.3 KiB Raw Blame History