- improve performance (fast as TBB - if possible) - lockfree-fifo, -lifo and -priority queue - check work-stealing algorithm
Maybe there are some things to work out before implementing lockfree structures. What would be helpful to know is : - task latency - you send one task in the system and see how long it takes (no op task) - throughput - how many tasks per second can you handle (no op task) - a combination of the two - running the same tests with tasks of arbitrary complexity A profiler running during these tests should give you a pretty clear picture about what needs to be improved. On my side, what I could do is profile the basic parallel_sort implementation I posted a while ago. Regards. -- Edouard Alligand Partner http://www.bureau14.fr/