Hosted forges' cores and the xlarge size (#13749)
A hosted forge's CPUs are its own: set apart from the work production does, and a size above large for a forge whose crew builds something as big as the forge.
FORGE_CHILD_CORES
An operator sets FORGE_CHILD_CORES to a core list such as 40-55. Unset (the
default) nothing changes: cells rotate over the cores the pod may use.
With it set, only a hosted forge's cell (its label starts forge:) is pinned,
by taskset, to a block of the listed cores chosen by its name alone, so a
release or a restart puts it on the same cores. A check's cell (check:), an
app's (app:) and any unlabelled cell are never pinned into the list.
- A size of
nCPUs takes a block ofncores; the block is one of
floor(pool / n) aligned blocks, so a pool that n does not divide leaves
its last cores unused.
- A size as big as the pool, or bigger, gets the whole pool.
- Two forges whose names hash to the same block share those cores: the
reservation keeps production off them, it does not give each forge its own.
- Cores outside the list are for production's own work. Keep them out of the
host's cpuset for production yourself; the forge only pins children.
- Cores the pod may not run on are never named.
xlarge
xlarge is 16 CPUs, 32 GiB of memory, 80 GiB of disk. It is chosen, on the new
forge screen or with size, and never stepped up to on its own: a forge that
ran out of memory is still given at most large's room. Moving a forge whose
disk holds 40 to 80 GiB gives it xlarge.