GPU Sharing & Fractional Allocation (E3)¶

Next-Gen 40 · feature E3 · ADR 0030 · spec design/vision/specs/E3-gpu-sharing.md

E3 makes fractional GPU allocation first-class across the scheduler abstraction and the K8s serving path. A workload asks for a fraction of a GPU (or a named MIG profile); the platform picks the best available mechanism for the cluster's capabilities, bin-packs fractional requests onto whole GPUs, surfaces the isolation level honestly, and accounts fractional GPU-hours.

Mechanisms (strongest → weakest isolation)¶

Mechanism	Isolation	Notes
mig	hardware	NVIDIA Multi-Instance GPU — memory/SM partitioned
timeslice	soft	time-sliced sharing — no memory/SM partition
whole	exclusive	a full GPU (fraction ≥ 1 or fallback)

Honest fallback¶

If a sub-1.0 fraction is requested but the cluster supports no fractional mechanism, the request is rounded up to a whole GPU and the wasted capacity is surfaced — the platform never pretends a share happened when it didn't:

exa hpc gpu-share plan JPCP --fraction 0.25
# JPCP: whole (isolation: exclusive) — 1.00 GPU
#   no fractional support — rounded 0.25 up to a whole GPU (75% wasted)
#   ⚠ 75% GPU capacity wasted by this fallback.

With MIG or time-slicing available:

exa hpc gpu-share plan JPCP --fraction 0.25 --mig-capable      # → MIG (hardware isolation)
exa hpc gpu-share plan JPCP --fraction 0.5 --timeslice         # → timeslice (SOFT isolation)
exa hpc gpu-share plan JPCP --mig 2g.10gb --mig-capable        # → exact MIG profile

Bin-packing¶

exa hpc gpu-share pack --ask a:0.5 --ask b:0.3 --ask c:0.4 --gpus 2 --timeslice
# packs a+? onto GPU 0 and the rest onto GPU 1, reporting GPUs used and any unplaced asks

First-fit-decreasing packing keeps GPU count low; asks that don't fit are reported as unplaced rather than silently dropped. Mechanisms are never mixed on one physical GPU.

Fractional accounting¶

exa hpc gpu-share plan JPCP --fraction 0.25 --timeslice --record
exa hpc gpu-share accounting

fractional_gpu_hours(fraction, seconds) = fraction × wall-hours, so a 0.25-GPU job for 2 hours books 0.5 GPU-hours — feeding the FinOps / carbon accounting (exa finops).

Programmatic use¶

from examlops.gpu_sharing import FractionalAsk, ClusterGpuCaps, select_mechanism, bin_pack

caps = ClusterGpuCaps(supports_mig=True, mig_profiles=["1g.5gb", "2g.10gb"])
choice = select_mechanism(FractionalAsk("JPCP", fraction=0.2), caps)
print(choice.mechanism, choice.isolation, choice.wasted_fraction)

result = bin_pack([FractionalAsk("a", 0.5), FractionalAsk("b", 0.4)], gpu_count=1, caps=caps)
print(result.gpus_used, result.unplaced)

Graceful degradation¶

All planning and accounting is pure-Python — no GPU or driver required. Capability flags come from HPC discovery (exa hpc detect/gpus); absent them, the honest-fallback path keeps allocations correct (whole GPUs, waste surfaced).