Hi Tom, On Thu, 20 Jun 2024 at 17:15, Tom Rini <tr...@konsulko.com> wrote: > > On Thu, Jun 20, 2024 at 05:05:26PM -0600, Simon Glass wrote: > > Hi Tom, > > > > On Thu, 20 Jun 2024 at 08:39, Tom Rini <tr...@konsulko.com> wrote: > > > > > > On Thu, Jun 20, 2024 at 07:33:46AM -0600, Simon Glass wrote: > > > > > > > Currently the world builds run on all runners, including faster and > > > > slower ones. > > > > > > > > The difference can be quite dramatic, with some builders 4x as fast as > > > > others, resulting in just one world build taking between 20 minutes and > > > > an hour and 20 minutes. > > > > > > > > Add a tag so that we can select which builders run these CPU-intensive > > > > jobs. > > > > > > > > With this tag we can also increase CPU utilisation by running multiple > > > > QEMU tests in parallel. Currently these tests leave most machines fairly > > > > idle, since we cannot run more than one world build on a machine. > > > > > > > > Signed-off-by: Simon Glass <s...@chromium.org> > > > > > > This conflicts I think with Jiaxun's desire to make our GitLab job > > > runnable on the public runners too, and where we'll end up with 10 world > > > build jobs ala Azure. > > > > It probably doesn't actually conflict, although I am not sure if one > > can add a tag to jobs that run on public runners. > > I mean conceptually at least as it will likely be slower to build the > world as 10 jobs than as 4 jobs.
Yes I've noticed that the per-board build time reduces as the number of boards increases. For example moa varies between about 0.53 builds per second (for ~100 boards), to 0.59 (for ~400 or ~600). But there are 64 threads, so we need at least a few hundred to get to a steady state. Regards, Simon > > -- > Tom