Usage of deprecated and removed tbb::task_scheduler_init and task API (Intel oneTBB)

[valpackett]

tbb::task_scheduler_init has been deprecated since Intel TBB 4.3 Update 5 and in the new Intel oneTBB 2021.x releases it is finally completely gone.

According to documentation,

Using task_scheduler_init is optional in Intel® Threading Building Blocks (Intel® TBB) 2.2. By default, Intel TBB 2.2 automatically creates a task scheduler the first time that a thread uses task scheduling services and destroys it when the last such thread exits.

and Intel does not recommend specifying the number of threads anywhere:

The reason for not specifying the number of threads in production code is that in a large software project, there is no way for various components to know how many threads would be optimal for other threads. Hardware threads are a shared global resource. It is best to leave the decision of how many threads to use to the task scheduler.

Please remove the usage of tbb::task_scheduler_init and the tbb/task_scheduler_init.h header, or at least detect their existence in the build system if you want to preserve the behavior with older TBB versions.

SuiteSparse/SPQR/Source/spqr_parallel.cpp

Lines 91 to 92 in 1869379

	task_scheduler_init
	init (nthreads <= 0 ? (task_scheduler_init::automatic) : nthreads) ;

---

Actually the whole tbb::task API is gone. task_group is suggested as a replacement.

[DrTimothyAldenDavis]

Thanks for the heads up ... that's a very unpleasant surprise. This is a serious issue because controlling the # of threads is a feature that I require to satisfy the needs of MATLAB. SPQR must follow the # of threads set by the MATLAB maxNumCompThreads (see https://www.mathworks.com/help/matlab/ref/maxnumcompthreads.html). TBB seems to be disallowing any control over the # of threads used. This function is used in x=A\b inside MATLAB, and this deprecation breaks it.

Intel's recommendation is an overstatement. I need to limit the task schedule in ways that no task scheduler inside TBB can predict, unless it can somehow glean the maxNumCompThreads itself from MATLAB. I'm happy to let TBB create its own tasks and threads but I have to have control over an upper bound on the number of threads it is permitted to use. It would be free to use any number of threads it likes, between 1 and my provided upper bound. But TBB has no mechanism for this. At least none that I can find.

With this deprecation and removal of task_schedule_init, TBB has become unusable to me. There seems to be no workaround. I will see about transitioning SuiteSparseQR to OpenMP tasking, and dropping the use of TBB entirely. I only use TBB within this single file, to execute a simple task tree. It shouldn't be too hard to switch to OpenMP tasks instead.

If there's a method for providing a max upper bound on the # of threads TBB is permitted to use, can you point it out to me?

[DrTimothyAldenDavis]

You'll note in my code that default is to let TBB pick the # of threads to use. It's in the non-default case that I require control.

[valpackett]

providing a max upper bound on the # of threads TBB is permitted to use

global_control?

[DrTimothyAldenDavis]

global_control would be fine in MATLAB, since it's an application, but not inside a library like SPQR. Intel doesn't recommend global_control inside a library.

I need to control the # of threads used when I combine my parallel library with the parallel BLAS. For that, I need specific control, such as mkl_set_num_threads_local ( https://software.intel.com/content/www/us/en/develop/documentation/onemkl-developer-reference-c/top/support-functions/threading-control/mkl-set-num-threads-local.html ).

global_control would be a sledge hammer, compared to a finer control like that function. So I don't think I can use global_control.

[Faroito]

Maybe you can use task_arena?

[DrTimothyAldenDavis]

It's not safe to switch to another feature of TBB. This is not a minor change to TBB.   I was using a core part of TBB, and now all features I was using in TBB are entirely gone. If Intel is willing to deprecate x=A\b in MATLAB, there's no guarantee task_arena will be around in 10 years either.

The solution is to switch to OpenMP tasking instead, and jettison all uses of TBB. 

[anton-potapov]

@DrTimothyAldenDavis it was 15 years from first version of TBB till the revamp (oneTBB, which removed tbb_scheduller_init). While it is not a guarantee, but IMHO it is safe to assume that task_arena will be there for 10 years :)

[DrTimothyAldenDavis]

Fixed in the draft SuiteSparse v5.11.0 (TBB is removed).