Eduardo Arango Gutierrez
Senior Systems Software Engineer @NVIDIA
Landsberg am Lech, Germany
Actions
Eduardo is a Senior Systems Software Engineer at NVIDIA, working on the Cloud Native Technologies team. Eduardo has focused on enabling users to build and deploy containers on distributed environments.
Area of Expertise
The Hidden Cost of Fat CRs: How NFD Slimmed Its Control-Plane Footprint
Node Feature Discovery is widely deployed across Kubernetes to turn hardware and system capabilities into labels consumed by schedulers, operators, and platform teams. As data centers add more devices and expose more features per node, NFD must publish more data, more often, across more nodes. That creates a multiplicative control-plane cost: nodes × object size × churn. We show how we profiled that cost and redesigned the path: moving built-in label computation to nfd-worker, publishing the resulting labels in NodeFeature.spec.labels, and adding noPublishFeatures to omit raw features. The result: 50% lower nfd-master memory use and an 80% smaller NFD footprint in etcd, with headroom that grows with node count. We also explain why cluster-scoped NodeFeatureRules remain centrally evaluated and extract reusable patterns for building scalable Kubernetes controllers.
The Hidden Cost of Fat CRs: How NFD Slimmed Its Control-Plane Footprint
Node Feature Discovery is widely deployed across Kubernetes to turn hardware and system capabilities into labels consumed by schedulers, operators, and platform teams. As data centers add more devices and expose more features per node, NFD must publish more data, more often, across more nodes. That creates a multiplicative control-plane cost: nodes × object size × churn. We show how we profiled that cost and redesigned the path: moving built-in label computation to nfd-worker, publishing the resulting labels in NodeFeature.spec.labels, and adding noPublishFeatures to omit raw features. The result: 50% lower nfd-master memory use and an 80% smaller NFD footprint in etcd, with headroom that grows with node count. We also explain why cluster-scoped NodeFeatureRules remain centrally evaluated and extract reusable patterns for building scalable Kubernetes controllers.
Operating Autonomous Agents in Production: An Open Reference Architecture for Sandboxing
The Q1 2026 OpenClaw security crisis, involving 30,000+ exposed gateways, demonstrated that agents frequently handle sensitive enterprise and customer data without adequate safeguards. For autonomous systems, security posture is not an optional layer; it is as fundamental as the underlying infrastructure.
This incident wave highlighted how easily ambient authority and flat networks become high-risk avenues for malicious exploitation. The lesson is not "don't run agents", it is that ambient authority, flat networks, and unsandboxed execution do not survive contact with autonomous software.
This talk presents a fully open-source pattern for providing a forward-deployed engineering team with always-on agents without granting them production access
At the end of this talk, the audience will walk away with a fully open-source architecture to safely deploy OpenClaw-based agents to serve a development team.
Running MCP Servers in Production: The MCP Lifecycle Operator
The Model Context Protocol (MCP) is emerging as the standard for connecting AI applications to external tools and data sources. But deploying MCP servers at scale on Kubernetes requires hand-wiring Deployments, Services, config rollouts, and health checks for each server with no standard API or ecosystem integration point.
The MCP Lifecycle Operator, a k8s-SIG Apps project, introduces a declarative MCPServer custom resource that automates the full server lifecycle: deployment, networking, configuration management, automatic rollouts on config changes, MCP-aware readiness via protocol handshakes, and security-hardened defaults.
In this session, we cover the operator's architecture and capabilities, demo deploying MCP servers on k8s, and discuss the project's roadmap for ecosystem integration — how MCP gateways, catalogs, and registries can build on the MCPServer primitive to deliver auth, traffic management, and discovery as a complete k8s-native solution for MCP infrastructure.
Fake It Until You Make It: Mocking NVML to Test the Kubernetes GPU Stack on Zero GPUs
GPU hardware is scarce, expensive, and hostile to fast CI/CD. Yet every component in the NVIDIA Kubernetes GPU stack, device plugins, DRA drivers, GPU Operator, depends on NVML, the user-space C library that talks to the driver. What if you could swap it for a drop-in mock that makes nvidia-smi, and any Go binary linked against go-nvml, believe it runs on a DGX A100, on your laptop, in a kind cluster, or in a GitHub Actions runner with zero GPUs?
This talk introduces nvml-mock, an open-source, YAML-configurable mock of libnvidia-ml.so implementing 89 NVML entry points via an auto-generated CGo bridge. We walk the architecture, C shim, a singleton Go engine with handle tables and reference-counted Init/Shutdown, and YAML device profiles for A100, GB200, and custom topologies, then show the same production binary tested end-to-end against simulated multi-GPU nodes. Finally, we show how to build your own GPU-aware CI pipelines without provisioning a single GPU.
From old-school HPC to sbatch: regulated industries using immutable systems & confidential computing
Getting a GPU cluster from racked to running distributed training is where weeks disappear. A specialised AI-cluster recipe can carry 268 config values across 16 components, and a single mismatch can cost one or two percentage points of training throughput. We treat that recipe as an artifact you capture, version-lock, and replay, then prove it live. On stage, on a bare-metal H100 cluster, we bootstrap baseline Kubernetes, snapshot it with NVIDIA AI Cluster Runtime (AICR) into a version-locked recipe, render it as Helm values for Spectro Cloud's PaletteAI (a declarative Cluster API platform), deploy Slurm via the Slinky slurm-operator, and submit a multinode sbatch job, racked to running, end to end.
You leave with a workflow you can replay on any conformant Kubernetes cluster, validated against the CNCF Kubernetes AI Conformance requirements, and AI/HPC defaults you'd otherwise learn the hard way. For AI-infrastructure maintainers bridging Slurm-native HPC and cloud-native AI.
Fake It Until You Make It: Contributing to the Kubernetes GPU Stack on Zero GPUs
GPU hardware is scarce and queued behind production, yet every component in the Kubernetes GPU stack (device plugins, DRA drivers, the GPU Operator) depends on NVML, the user-space library that talks to the driver. Owning a GPU has therefore been a prerequisite for contributing to the layer. nvml-mock removes it: an open-source, YAML-configurable fake of libnvidia-ml.so that makes nvidia-smi and any go-nvml binary behave as if running on a DGX A100 or GB200 NVL node, whether on a laptop, in kind, or in a GitHub Actions runner with no GPUs attached. We describe what changed once contributors could run end-to-end tests without hardware: reproducible reviews, regression gates in CI, and a feedback loop no longer blocked on a hardware queue. We also cover the transferable pattern for other accelerator vendors, mocking the vendor library behind a CGo bridge with declarative device profiles, increasingly relevant as WG Device Management drives DRA to GA.
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top