Maximum task number

Maximum task number

Hi!

Is any restrictions on creating task number, maximum task number?

6 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.
Best Reply

Performance will suffer if you don't practice "recursive parallelism", where every task generates only (roughly) the number of tasks necessary to create "parallel slack" that allows the scheduler to efficiently divide the work for execution with efficient use of the local cache (both aspects are important). Otherwise, feel free to fill physical RAM and even your swap space with tasks before you even start executing them.

In following up on Raf's comment...
Assume you have a large input dataset in a file that can be divided up into a list of tasks. You could program (pseudo code)

while read chunk
parallel task chunk
end while

The above could concievably enqueue as many task as there are chunks in your data set (chunk == task sized hunk of data).
More importantly, it is concievable (though not probable) that the read loop could complete before the first enqueued task completes.
IOW memory filled (or overflowed) with tasks.

A better strategy in the above situation is to use the parallel_pipeline or similar strategy, where the read loop (pipe in pipeline) is dependent upon the availability of a buffer (token). Using this method limits the number of pending task to a good (efficient) working set.

Jim Dempsey

www.quickthreadprogramming.com

The answer can be given at different levels of detail, of course, and maybe even my warning against the use of large numbers of simultaneous tasks was superfluous. It is always nice when a motivation is given together with a question, so that we don't have to guess what level of detail would be useful, etc.

More technically, I am not aware of any relevant artificial restriction, either at any one time, or, I might add, over the lifetime of a program. I remember a problem with an algorithm whose implementation was susceptible to a wraparound problem, but that was above the task level and has since been remedied.

Thanks, guys!

Hi everybody,

Daniil,

Raf and Jim already answered your question. I'd like to provide some sources. Unfortunately, I don't have a stress test for 'tbb_task' class but I've enclosed a 'tbbthreadtest.txt' file with a stress test for 'tbb_thread' class. Take a look if interested.

Best regards,
Sergey

Fichiers joints: 

Fichier attachéTaille
Télécharger tbbthreadtest.txt4.61 Ko

Connectez-vous pour laisser un commentaire.