per-thread data persisting through parallel task asignment

[Brian_Budge]

Hi all -

I hope someone can help me with this issue. I'm using the tbb::task library to perform a parallel post-order traversal. The problem I'm facing is that I need to generate random numbers as part of my traversal process. I already have a class to perform the calculations (SFMT derived), and I've used it in threaded programming before. The problem is that when using tbb::task, there is no access to a per-thread memory (at least not that I've found). Any ideas?

Thanks!

[Brian_Budge]

I think what would be ideal for me from a performance standpoint would be to have an array of N (N is # running threads) SFMT states which can be referenced in the task. Then if I could get a thread index during traversal, I could use the appropriate SFMT state for that thread. Can this be done?

[Alexey-Kukanov]

Quoting - Brian Budge

Hi all -

I hope someone can help me with this issue. I'm using the tbb::task library to perform a parallel post-order traversal. The problem I'm facing is that I need to generate random numbers as part of my traversal process. I already have a class to perform the calculations (SFMT derived), and I've used it in threaded programming before. The problem is that when using tbb::task, there is no access to a per-thread memory (at least not that I've found). Any ideas?

Thanks!

Look at enumerable_thread_specific and combinable classes in TBB 2.2; they are designed to help in cases like yours.

[Brian_Budge]

Thanks for the answer Alexey. In the past when I've used tls with boost threads, the overhead was very large. Should I expect much overhead when using the enumerable tls data in tbb?

[Bartlomiej]

Quoting - Brian Budge

Thanks for the answer Alexey. In the past when I've used tls with boost threads, the overhead was very large. Should I expect much overhead when using the enumerable tls data in tbb?

I did only one program with enumerable_thread_specific, up to now, but there was no noticable difference with respect to compiler-specific API for the TLS. technically there must be some overhead, but it seems negligible.