[starpu-devel] Trouble getting performance from a simple program

[Troels Henriksen <athas@sigkill.dk>]

Hello starpu-devel, I am trying to understand how to write hybrid
applications. Specifically, I have adapted the mandelbrot example from
the STarPU distribution, taken out all the graphics stuff, and added
simplistic instrumentation. I want to start by understanding
partitioning of computation (with only trivial data).

This is my program so far:
https://github.com/diku-dk/hybrid-starpu/blob/main/mandelbrot/mandelbrot.c

However, I find that my program runs very slowly on a machine with
2xA100 GPUs and two 24-core CPUs. Specifically, using nvtop I see that
the program utilises both the CPUs and the GPUs, but that the GPUs
finish the work very quickly (as expected), and a lot of time is then
spent waiting on the CPUs. I have played around with increasing the
number of blocks (for overpartitioning), but it doesn't seem to make a
significant difference. So questions are:

1) Where can I read up on the basics of load balancing in StarPU?

2) Is there already an example of an efficient Mandelbrot program (or similar
   compute-bound program) I can study?