Re: [starpu-devel] Follow-Up on Data Handle Scope and Task Scheduling

[Xinbo Li <lix34545@myumanitoba.ca>]

Hi Samuel, 

Thanks for the clarification. Just to make sure I fully understand your point regarding "you'd prefer to just let starpu handle allocation by passing a NULL pointer (and possibly use a task to initialize data in
it)", do you mean that, instead of the following approach:

// Method 1. the developer allocates memory for data then register to StarPU

      DataType data;

      preparation(&data);

      starpu_data_handle_t handle;    

      starpu_custom_data_register(&handle, STARPU_MAIN_RAM, &data);

It is more recommended to do:

// Method 2. StarPU handles the memory allocation for data

      starpu_data_handle_t handle;    

      starpu_custom_data_register(&handle, -1, nullptr);

      starpu_task_insert(&codelet_that_prepares_data, STARPU_RW, handle, 0);

Please let me know if I’ve captured your suggestion correctly.

Best,
Xinbo

---

From: Samuel Thibault
Sent: Thursday, April 17, 2025 5:47 PM
To: Xinbo Li
Subject: Re: Follow-Up on Data Handle Scope and Task Scheduling

Caution! This message was sent from outside the University of Manitoba.


Xinbo Li, le ven. 18 avril 2025 00:44:39 +0000, a ecrit:
> In that case, if we define a data handle inside a local block and unregister it
> immediately after task submission, any subsequent tasks — even if unrelated —
> would be forced to wait for the earlier tasks to complete, due to the
> synchronization imposed by the unregister operation. This could potentially
> reduce efficiency.

Yes. That's why you'd prefer to just let starpu handle allocation by
passing a NULL pointer (and possibly use a task to initialize data in
it), and use starpu_data_unregister_submit after inserting all tasks
using it.

>     // Task 1
>     starpu_task_insert(
>         &some_codelet,
>         STARPU_RW, handle,
>         0);
>
>     starpu_data_unregister(handle); // unregister at end of block
> }
>
> // Task 2
> starpu_task_insert(
>     &some_other_codelet,
>     STARPU_RW, some_other_handle,
>     0);
>
> In this case, would Task 2 be forced to wait for Task 1 to finish,

starpu_data_unregister waits for the completion of task 1, yes.

Samuel