Go from a single pod to petaflop scale: run large-scale AI and HPC on Kubernetes with Kueue, Ray, Karpenter, JobSet, GPUs, and real industry workloads.
Key Features
- Build the whole stack, from a single pod to supercompute scale
- Migrate SLURM, PBS, and LSF batch workloads to Kubernetes-native job patterns with deployable, production-tested examples
- Apply the stack to real finance, genomics, engineering, and large-scale AI and LLM workloads through hands-on labs and case studies
Book Description
Simulations, training, and inference once ran on separate clusters. Now they share GPUs, networks, and schedulers, and Kubernetes is becoming the platform that brings them together. From Pods to Petaflops treats cloud-native HPC and AI as one end-to-end discipline, taking you from a single kubectl apply to compute-intensive workloads at supercomputing scale. Written by Kubernetes practitioners Sanjeev Ganjihal and Mesut Özdil, the book replaces scattered blogs, READMEs, and conference talks with a practical path to production. You'll map familiar HPC concepts such as SLURM jobs, queues, MPI, schedulers, and shared storage to Kubernetes, then build platforms with Kueue, MultiKueue, JobSet, LeaderWorkerSet, Ray, Karpenter, Cluster Autoscaler, and NVIDIA GPU tooling including MIG, DRA, DCGM, and HAMi. Hands-on examples cover Monte Carlo risk, genomics, CFD, weather, seismic modeling, engineering simulation, distributed training, and multi-host LLM inference. You'll also tackle observability, FinOps, security, multi-tenancy, multi-cluster scheduling, and the challenges that emerge at 10,000 nodes. By the end, you'll be able to design, deploy, operate, and optimize Kubernetes-based HPC and AI platforms, from a single pod to petaflop-class production workloads.
What you will learn
- Map SLURM, PBS, and LSF concepts to Kubernetes primitives
- Schedule batch jobs with Kueue, MultiKueue, JobSet, and LeaderWorkerSet
- Run distributed Python and AI workloads with Ray on Kubernetes
- Autoscale CPU and GPU nodes with Karpenter and Cluster Autoscaler
- Manage and share GPUs with MIG, DRA, DCGM, and HAMi
- Design high-throughput storage and checkpoint/restart for HPC and AI
- Build finance, genomics, engineering, and AI-training platforms
- Operate secure, observable, cost-aware multi-tenant clusters at scale
Who this book is for
This book is for HPC engineers, platform and DevOps engineers, SREs, Kubernetes engineers, data scientists, and technical leaders who need to run batch, distributed, GPU, or AI workloads on Kubernetes. Familiarity with either Kubernetes or traditional HPC systems is helpful, along with basic Linux, container, and YAML knowledge.
