Files
werewulf/hpcod/examples/gpu-request.yaml
T
ilker d39c26d39d hpcod phase 2: placement
placement.Plan(sites, request, weights) is a pure function - no I/O, no cloud -
which is what makes the decision reproducible, unit-testable against synthetic
multi-site fixtures, and usable as a dry-run endpoint later. hpcod plan can
run against live sites or a saved discover -json snapshot.

Filter then score. Filtering eliminates the impossible and records why for
each site; scoring ranks the survivors with capability_fit weighted highest,
because scarce hardware must be kept for requests that need it. A plain CPU
job placed on the only InfiniBand site is a scheduling failure even though it
fits, and the tests assert exactly that.

The refusals carry the design's central distinction. A site is told apart as
"none present", "present but unconfirmed - the fabric has not been verified by
a node agent", and "present but not passthrough-safe (shared IOMMU group)".
Those are three different problems needing three different responses: buy
hardware, run the agent, fix the topology. Reporting them all as "no capacity"
would waste an operator's afternoon.

Requests state usable memory; the portal adds the in-RAM rootfs reserve on
top, since a tenant allocated exactly what it asked for would OOM under its
own job. Packing prefers hosts already in use by the tenant so whole
hypervisors stay free for later requests.

Verified against real hv01 inventory, where the request for 4 GPUs and
InfiniBand is refused with "gpu: need 4, only 1 usable" and "infiniband: none
present" - correct, because the ConnectX-4 Lx cards are RoCE and the curated
map knows it.

Also: an invalid lease now names the field instead of surfacing a bare
"invalid duration", which a test caught.
2026-09-01 12:08:26 +02:00

27 lines
449 B
YAML

# A tenant cluster that needs GPUs and a real fabric.
name: genomics-week12
lease: 14d
head:
cores: 4
memory_gb: 16
disk_gb: 200
compute:
count: 4
cores: 8
memory_gb: 32 # usable; the in-RAM rootfs reserve is added on top
require:
- capability: gpu
count: 1
vendor: nvidia
- capability: infiniband
count: 1
prefer:
- capability: nvme
stack:
os: rocky9
scheduler: slurm
identity: freeipa