Lesson 1C - GPU Sharing & Fractional GPUs with HAMi¶

Get the code to run this lab

The commands on this page come from the repository, not the website. Clone it and enter this lesson's folder: git clone https://github.com/ld-singh/ai-factory-ops-lab && cd ai-factory-ops-lab/portfolio-lab/01-k8s-gpu-platform/hami. Browse this lesson on GitHub

Course home: AI Factory Operations Lab · Previous: Lesson 1B - KAI Scheduler · Next: Scheduling simulation lab (the hands-on part of this lesson)

🧪 Want to run it? Jump straight to the hands-on labs: ▶ Scheduling simulation (free, no GPU) · ▶ Isolation on a real GPU (validated on an RTX A6000). The rest of this page is the concepts behind them.

So far you've requested a GPU as an indivisible integer: a pod asks for nvidia.com/gpu: 1 and gets a whole device. Lesson 1's default scheduler can do nothing else, and Lesson 1B's KAI exercises kept to whole GPUs on purpose - to focus on queue policy (KAI can itself slice GPUs, via a gpu-fraction request; see the contrast below). In real fleets, whole-GPU allocation is often wasteful - an inference pod using 3 GB of an 80 GB A100 strands the other 77 GB. GPU sharing is how platforms fix that, and HAMi (Heterogeneous AI Computing Virtualization Middleware, a CNCF sandbox project) is an open-source way to do it: it lets one physical GPU be requested as fractions - "2 GB of GPU memory and 30% of compute" - with the scheduler packing multiple pods onto one device and a runtime layer enforcing the limits inside each container.

This lesson matters doubly for this course's mission of learning cheaply:

The sharing decision is control-plane logic - like queueing in Lesson 1B, the bin-packing of fractional requests onto devices is studyable without hardware.
On real hardware, sharing is a cost multiplier. When you rent one GPU for Lesson 6, HAMi turns it into several - you can run multi-tenant sharing, oversubscription, and isolation experiments on a single cheap VM that would otherwise need a fleet.

🎯 Learning objectives - after this lesson you can:

Explain the four mainstream GPU-sharing mechanisms - time-slicing, MPS, MIG, and software virtualization (HAMi) - and the isolation guarantee each does and doesn't give.
Explain how HAMi's pieces fit: mutating webhook → scheduler (extender) → device plugin → in-container enforcement (HAMi-core intercepting CUDA calls).
Install HAMi on a real-GPU node (the Lesson 6 machine) and run multiple pods sharing one physical GPU, each seeing only its memory slice in nvidia-smi.
Demonstrate that fractional scheduling arithmetic (does a 0.5-GPU request fit?) is a control-plane decision, while memory/compute isolation is runtime enforcement that only real hardware can prove.
Say precisely which sharing claims belong on which side of the fake/real line.

🧭 Mode: 🟦 Simulation. The concepts and the runnable scheduling sim are free and no-GPU. The real isolation half is executed as part of Lesson 6 - Real GPU in your one rental session - it is described here, next to the concepts it proves, but run there.

📋 Prerequisites: Lesson 1 (mental model). The real isolation half needs the Lesson 6 host set up (driver, toolkit, Kubernetes) - all part of that one rental session.

🧪 Two runnable labs in this lesson¶

The Parts below are the concepts. The hands-on work ships as two self-contained labs on opposite sides of the sim/real line - each with its own Makefile and pinned versions:

🟦 Scheduling simulation - no GPU, free. Prove HAMi's scheduling decisions on a fake GPU fleet, on your laptop: a fractional request is placed, an over-large request is rejected (Pending + reason), and the per-pod placement decision (FilteringSucceed) is visible. GPU sharing (multiple pods co-resident on one device) and isolation are not here - the fake fleet can't do them without forcing - so they live in the real-GPU half below. Free.

🟥 Isolation on a real GPU - the real half, run in Lesson 6. Two pods sharing one physical card, a virtualized nvidia-smi, and a CUDA allocation refused at the slice limit - the half a simulation can never prove. It runs on one cheap GPU as part of the Lesson 6 one-rental session, but it's documented here next to the concepts it proves.

Do the simulation now, for free; the real isolation lab is waiting for you in Lesson 6.

⚠️ ILLUSTRATIVE-manifest rule (same as Lesson 1B): HAMi is actively developed. Chart values, resource names, annotations, and defaults can change between releases. Snippets below show the shape; confirm exact names against the official docs before running: https://project-hami.io/ and https://github.com/Project-HAMi/HAMi. The manifests you actually run, and their output, go into ../../06-validation-reports/.

One physical GPU, many workloads. Four mainstream ways to slice it:

Mechanism	How it works	Memory isolation	Compute isolation	Needs special hardware?
Time-slicing (NVIDIA device plugin option)	Advertise N "replicas" of each GPU; contexts take turns	❌ None - any pod can OOM the device for everyone	❌ None (round-robin, no limit)	No
MPS (Multi-Process Service)	CUDA contexts share the GPU concurrently via a daemon	⚠️ Limited (resource limits per client, weaker fault isolation)	⚠️ Partial (active thread percentage)	No
MIG (Multi-Instance GPU)	Hardware partitions the GPU into isolated instances	✅ Hardware-enforced	✅ Hardware-enforced	Yes - only specific datacenter cards: A100/A30, H100/H200, Blackwell. Not Ada Lovelace (L4/L40S)
HAMi (software virtualization)	Device plugin advertises virtual GPUs; HAMi-core library intercepts CUDA calls in-container to cap memory/compute	✅ Software-enforced (CUDA-API level)	⚠️ Software-enforced (core-percentage throttling)	No - works on consumer and datacenter GPUs

💡 Why HAMi is the one worth learning here¶

Each of the other three mechanisms fails this course's test in a different way:

MIG needs an expensive A100/H100-class card, and it partitions in hardware - something you enable, not something you operate.

Time-slicing gives no isolation, so there's nothing to prove.

MPS only partially isolates.

HAMi gives enforceable memory limits on a GPU you can actually afford to rent - a T4 or RTX-class card - and it is the one mechanism that surfaces both halves of this course's central distinction, in a single YAML block:

The two halves What it is Where this course proves it

① Scheduling decision which device gets which fraction free, on fake GPUs - hami-scheduling-sim/

② Runtime enforcement the cap held inside the container one cheap real GPU - hami-isolation-realgpu/

Master that line - decision vs enforcement - and you can evaluate any GPU-sharing technology, not just HAMi. It is the single most transferable idea in this lesson.

And it isn't a niche bet. In June 2026 NVIDIA's own KAI Scheduler adopted HAMi-core as the engine that enforces its fractional-GPU memory limits (PR #60). So the runtime layer you learn here is the same one a flagship scheduler now leans on - HAMi-core is becoming the de-facto isolation layer, not an underdog.

How HAMi's pieces fit together:

pod requests nvidia.com/gpu + gpumem/gpucores fractions
        │
        ▼
mutating webhook        routes the pod to HAMi's scheduler
        │
        ▼
hami-scheduler          (kube-scheduler + extender) picks a node AND a specific
        │               device with enough remaining memory/core budget
        ▼
hami-device-plugin      advertises virtual GPUs, mounts HAMi-core into the container
        │
        ▼
HAMi-core (libvgpu.so)  intercepts CUDA driver API calls inside the container,
                        enforcing the memory cap and throttling compute

✅ Checkpoint (concept): without looking at the table, name the sharing mechanism you'd pick for (a) hard multi-tenant isolation on H100s, (b) squeezing three small inference services onto one rented T4, (c) quick-and-dirty test parallelism where isolation doesn't matter. (MIG / HAMi / time-slicing.)

The fractional resource model¶

With HAMi installed, a pod can ask for a slice instead of a whole device:

# ILLUSTRATIVE pod fragment - confirm resource names/units in the HAMi docs.
resources:
  limits:
    nvidia.com/gpu: 1          # number of (virtual) GPUs
    nvidia.com/gpumem: 2000    # GPU memory for this slice, in MiB
    nvidia.com/gpucores: 30    # % of the device's compute for this slice

💡 Read that the way the scheduler does. These are still just numbers - the scheduler's job is arithmetic: "device 0 has 16 384 MiB total, 6 000 MiB already promised, so a 2 000 MiB request fits." That bookkeeping is exactly the kind of control-plane decision Lessons 1 and 1B taught. What is new here, and not arithmetic, is what happens after placement: HAMi-core making a CUDA malloc beyond 2 000 MiB actually fail inside the container. Decision vs enforcement - the course's fake/real line, drawn through a single YAML block.

Contrast with KAI Scheduler, which also does GPU fractions (a gpu-fraction: 0.5 request, plus time-slicing/MPS sharing). The difference is emphasis, not capability: KAI's strength is queue policy (quota/borrow/reclaim/gang); HAMi's strength is per-device memory slicing with in-container enforcement. Real platforms combine a queue layer with a sharing layer.

Everything above is the decision side, and you study it for free. The other half - HAMi actually holding the cap inside a running container - is the one thing only a real GPU can prove.

On one cheap card, HAMi runs two pods on a single physical GPU, each capped at its own memory slice. Inside a capped pod, nvidia-smi reports only the slice (not the card's full memory), and a CUDA malloc past the cap fails even with plenty of memory free elsewhere on the device. That allocation failure is the single artifact that separates enforced isolation from bookkeeping.

🔬 Proved on real hardware: concurrent multi-pod sharing of one device, and software-enforced memory isolation at the CUDA API level. Not proved: hardware-grade isolation (that's MIG), performance interference under sustained load, or fleet-scale behaviour.

▶ This is a full, runnable lab - the install steps, manifests, pinned versions, and captured evidence live in Isolation on a real GPU. Run it during your one Lesson 6 rental (validated on an RTX A6000).

Two paired runnable lessons¶

The concepts above are split into two self-contained sub-lessons that sit on opposite sides of the scope boundary. Each has its own Makefile and pinned versions.

hami-scheduling-sim/ - control plane, no GPU. Based on the official HAMi local-fake-gpu tutorial: a kind cluster with real workers, the run.ai fake-gpu-operator advertising nvidia.com/gpu, HAMi with its device plugins off, and a node-registration annotation. Validates HAMi's scheduling decisions: a fractional request is placed, an over-large request stays Pending (CardInsufficientMemory), and the per-pod placement decision (FilteringSucceed) is visible. It does not do GPU sharing (multiple pods on one device) or isolation - the fake fleet can't without forcing - so those are the real-GPU lab below.
hami-isolation-realgpu/ - data plane, one cheap real GPU. Two pods share a single consumer 24 GB card; you observe the virtualized nvidia-smi and a CUDA allocation refused at the slice limit. Validates runtime isolation, which the simulation cannot.

The official tutorial notes gpumem/gpucores "require a real GPU"; that is about runtime isolation. Fractional scheduling works on fakes once HAMi's scheduler has the per-GPU memory/core figures, which the sim lesson supplies via the hami.io/node-nvidia-register annotation. The split keeps the boundary clear: the sim proves the placement decision, the real-GPU lesson proves enforcement.

Getting to production: most real clusters install the GPU stack with the NVIDIA GPU Operator, and running HAMi alongside it has one sharp edge (both ship a device plugin advertising nvidia.com/gpu). hami-gpu-operator-coexistence/ walks the clean single-node setup: the Operator manages the base stack (in production typically the driver and toolkit too; on the course VM those come from the image, so it runs with driver.enabled=false and toolkit.enabled=false), HAMi owns the device plugin, and it covers the reboot race that trips people up. Validated on an NVIDIA L40.

What you can and cannot learn here - the precise line¶

Capability	Where it's learnable	Why
Sharing-mechanism trade-offs (time-slicing/MPS/MIG/HAMi)	✅ Free (concepts)	Concepts
Fractional resource model & scheduling arithmetic	✅ Free (the sim lab; 1B for quota math)	Control-plane bookkeeping
Multi-pod co-residency on one device	🟥 Real GPU (isolation lab)	Needs a real device plugin path
Memory-cap enforcement inside the container	🟥 Real GPU (isolation lab)	HAMi-core intercepts real CUDA calls
Compute throttling accuracy / interference under load	❌ Out of scope	Needs sustained real workloads + measurement
MIG hardware partitioning	❌ Out of scope	Needs a MIG-capable card (A100/A30, H100/H200, Blackwell) - not Ada Lovelace

Operational takeaways¶

Sharing is two problems, not one: placement (scheduler) and enforcement (runtime). Evaluate any sharing solution by asking what enforces the limit and what happens when a tenant exceeds it.
No isolation = no multi-tenancy. Time-slicing is fine for your own test pods, unacceptable across teams: one OOM kills everyone's work on the device.
Fractionalization is the single biggest cost lever for inference fleets - small models on dedicated big GPUs are the most common money fire in AI infra.

➡️ Next: run the hands-on lab - HAMi scheduling simulation (free, no GPU). Then continue to Lesson 1D - GPU fleet scale simulation with Volcano for gang scheduling at fleet scale, or on to Lesson 2 - Slurm GPU Workload Management. The real isolation half of this lesson is run later, in Lesson 6 - Real GPU, alongside the other real-hardware work - one rental, everything at once.

The two halves	What it is	Where this course proves it
① Scheduling decision	which device gets which fraction	free, on fake GPUs - `hami-scheduling-sim/`
② Runtime enforcement	the cap held inside the container	one cheap real GPU - `hami-isolation-realgpu/`

Lesson 1C - GPU Sharing & Fractional GPUs with HAMi¶

🧪 Two runnable labs in this lesson¶

The GPU-sharing landscape¶

💡 Why HAMi is the one worth learning here¶

The fractional resource model¶

Sharing on a real GPU (the enforcement half)¶

Two paired runnable lessons¶

What you can and cannot learn here - the precise line¶

Operational takeaways¶