[Starpu-devel] StarPU task overhead

[Marcus Holm <marcus.holm@it.uu.se>]

Dear StarPU developers,

I'm having a little trouble understanding some numbers I'm getting  
from a small benchmark that I'm using.  It seems like creating and  
scheduling a task has an overhead of between 3 and 6 milliseconds,  
depending on the scheduler and compiler options.  That seems much too  
large, and I am at a loss to explain what is happening.  I have  
compiled with --enable-fast, without any debug options, and I don't  
believe the overhead is due to generating the performance model.  The  
benchmark is creating tasks with the same data usage pattern as your  
"matmul" example, but runs a sleep function instead of performing real  
work.  Your website shows task creation and scheduling overheads that  
are 3 orders of magnitude less, so I'm sure I must be doing something  
wrong. Can you please give me some advice?

Thanks and regards,
Marcus Holm