GPU Sharing on Kubernetes: DRA vs HAMi – Strategic Implications for North‑East India

Introduction

The rapid adoption of artificial intelligence (AI), high‑performance computing (HPC), and edge analytics has turned graphics processing units (GPUs) into a critical commodity for modern enterprises. In India’s North‑East region—home to a growing network of research institutes, start‑ups, and government‑backed innovation hubs—access to high‑end GPUs remains limited by both geography and budget. According to the Ministry of Electronics and Information Technology, only 12 % of the nation’s AI‑focused research labs are located in the North‑East, yet these labs account for roughly 8 % of the country’s total AI publications. The disparity underscores a pressing need for more efficient utilization of existing GPU assets.

Two competing approaches have emerged within the Kubernetes ecosystem to address this need: Dynamic Resource Allocation (DRA), a native extension to the Kubernetes scheduler, and the HAMi (Heterogeneous Accelerated Machine Interface) project, a community‑driven framework that predates DRA. Both aim to slice a single physical GPU into multiple logical partitions, allowing several containers to share the same hardware without sacrificing performance guarantees. This article dissects the technical evolution from opaque device counters to structured resource claims, evaluates the practical trade‑offs of DRA versus HAMi, and explores how each model can reshape the computational landscape of the North‑East.

Main Analysis

Historical Context: From Monolithic Device Plugins to Fine‑Grained Allocation

Early Kubernetes clusters relied on the device plugin API, which exposed a GPU as a single integer resource (e.g., nvidia.com/gpu: 1). This model forced a pod to claim an entire card, even if its workload required only a fraction of the device’s memory or compute capacity. The consequence was severe under‑utilization: a typical deep‑learning inference job might use less than 10 % of a 16 GB GPU, yet the scheduler would reserve the whole card, leaving the remaining 90 % idle.

To mitigate this inefficiency, the community introduced the HAMi project. HAMi layered a mutating webhook, a scheduler extender, and custom annotations on top of the existing device plugin. Users could specify memory (e.g., 8000Mi) and compute percentages (e.g., 10%) in pod specifications. The extender would then filter nodes, select a device UUID, and store the decision in a string annotation that only HAMi understood. While functional, this approach suffered from three major drawbacks:

• Opacity: The scheduler could not natively interpret the annotations, leading to sub‑optimal placement decisions.
• Complexity: Deploying HAMi required a bespoke webhook, an extender binary, and a custom runtime, increasing operational overhead.
• Fragmentation: Because the annotations were not part of the core Kubernetes API, other tools (e.g., autoscalers, monitoring agents) could not reliably query GPU usage.

Dynamic Resource Allocation (DRA): A Native Evolution

Recognizing the limitations of the HAMi stack, the Kubernetes SIG‑Node community introduced Dynamic Resource Allocation (DRA) in version 1.26. DRA extends the scheduler’s API to accept structured resource claims that describe not only the quantity of a resource but also its qualitative attributes (memory, compute, bandwidth). In practice, a pod can now request a gpu claim with a memory field of 8Gi and a core field of 0.2, and the scheduler will treat these as first‑class citizens during placement.

Key technical advantages of DRA include:

• Native Scheduler Integration: The scheduler evaluates claims directly, eliminating the need for external extenders.
• Standardized API: Claims are expressed in ResourceClaim objects, making them discoverable by any Kubernetes‑aware component.
• Fine‑Grained Accounting: The kubelet can enforce per‑container limits on memory and compute, reducing the risk of “noisy neighbor” interference.

Performance and Cost Implications

Empirical studies conducted by the Indian Institute of Technology (IIT) Guwahati in 2023 compared DRA‑enabled clusters against HAMi‑based clusters across three benchmark workloads: image classification (ResNet‑50), natural language processing (BERT‑base), and remote‑sensing inference (U-Net). The results were striking:

By improving average utilization from 42 % to 68 %, DRA reduced the effective cost per GPU‑hour by roughly 27 %. For a typical North‑East research lab that operates a 4‑GPU node at $2,500 per month, the savings translate to approximately $650 annually—money that can be redirected toward data acquisition or talent development.

Regional Impact: Enabling AI‑Driven Development in the North‑East

The North‑East’s topography—characterized by mountainous terrain, dense forests, and dispersed settlements—creates unique data challenges. Projects ranging from flood prediction in Assam to biodiversity monitoring in Arunachal Pradesh rely on processing large satellite imagery datasets. Historically, these initiatives have depended on cloud‑based GPU rentals, incurring costs that exceed $0.90 per GPU‑hour on average (according to a 2022 cloud‑provider price survey). By deploying on‑premise Kubernetes clusters with DRA, institutions can cut these expenses by up to 30 % while retaining data sovereignty, a critical factor for government‑sponsored research.

Moreover, the ability to share GPUs among multiple pods encourages collaborative workflows. A single GPU can now host a training job, an inference service, and a monitoring agent simultaneously, fostering a “single‑node AI hub” model that aligns with the Indian government’s “Digital India” vision for regional self‑reliance.

Examples

Case Study 1 – IIT Guwahati’s “Smart Forest” Initiative

In 2024, IIT Guwahati launched a pilot program to detect illegal logging using drone‑captured video streams. The team deployed a 2‑node Kubernetes cluster equipped with four NVIDIA A100 GPUs. By leveraging DRA, each GPU was partitioned into three logical slices: 0.4