For now, Colcon automatically pass -j{os.cpu_count()} -l{os.cpu_count} to the cmake --build command, unless the MAKEFLAGS environment variable contains them. There is no other way to control what arguments the builder receives. This works fine with make, but is useless with Ninja, which doesn't use an environment variable.

As a workaround, I use taskset to change the CPU affinity and limit the number of parallel build, but it only limit the number of parallelism and thus doesn't work with tools such as distcc which allow to build with more parallelism.

It would be great to at least be able to control the parallelism via a colcon flag, or perhaps better, be able to directly pass arguments to the cmake --build invocation.