tbb allocators

[Petros_Mamales]

Hi,
Currently whenever I want my threads to use memory, I typically allocate the necessary memory outside the tbb calls (parallel_for, say) and provide the thread-called functors with the appropriate information,effectively creating a -very- primitive heap private to my threads.
(Typically, this would be of importance, if I used single-threaded mkl linear equation solver where a number of auxiliary buffers would be requested for the solution.)
I was wondering if the tbb allocators do that for me automatically (even effectively), and therefore I do not need to do all this "maneuvre" - honestly, no attempt here to "fish" what the allocators do -)).
If not, how much of a penalty is incurred by using the tbb allocators? -I know they are to be faster than the os provided ones.
Thank you in advance for your help,
Petros

[RafSchietekat]

The scalable allocator does in fact have something like one heap per thread. Memory is allocated efficiently from the local "heap", and deallocating local memory is very fast compared to deallocating nonlocal memory.

[jimdempseyatthecove]

The TBB scalableallocator will generally be much faster than using the standard heap and near the performance of your own targeted memory pool. The potential downsides (caveats) are:

Memory consumption will increase (pool retention by thread).
There may be allocation/deallocation patterns in your application that aggravate the memory consumption.
Potential issues when static objects allocate and deallocate memory (before main).
Potential issues where a DLL is using a different thread pool (e.g. OpenMP) and your app is using TBB with overloaded new/delete.

Don't let this caviats interfere with your experimentation with using the TBBscalable allocator. Much thought has been put into the allocators to yield a superior solution (for most programming situations).

Jim Dempsey

[Petros_Mamales]

Raf,Jim,
Thank you very much for your response. Knowing this, lifts a big weight off my shoulders.
One final clarification: your comments on the tbb scalable allocator carry over to the cache aligned one as well?
Thank you for all your help,
Petros

[RafSchietekat]

Unless I missed something, the cache-aligned allocator is (still) based on the same code, so yes.

[Petros_Mamales]

Hi Raf,
Thank you for the confirmation.
Petros

[Petros_Mamales]

Hi Raf,
Thank you for the confirmation.
Petros