Setting Up Managed Lustre Storage on Google Cloud

Building Managed Lustre Storage on Google Cloud

Table of Contents

1. Overview

2. Excerpts from Official Documentation

3. Hands-on: Command Execution and Troubleshooting

4. Overview

1-1 What Is Google Cloud Managed Lustre?

Google Cloud Managed Lustre is a fully managed parallel file system engineered specifically for massive High Performance Computing (HPC) and AI/ML workloads. By offering the industry-standard Lustre file system (powered by DDN EXAScaler) as a cloud-native service, it delivers petabyte-scale storage and hundreds of GB/s of throughput out of the box—freeing you from the complexities of managing underlying storage infrastructure.

1-2 Google Cloud High-Performance Storage Options and When to Choose Lustre

Google Cloud provides multiple high-performance storage options tailored to the demands of different workloads:

Local SSD: Directly attached to VMs to deliver the lowest latency and highest IOPS, but the data is volatile (ephemeral) and cannot be shared across multiple VMs.

Managed Lustre: Distributes data across numerous storage servers to execute parallel I/O.

Filestore (High Scale/Enterprise): A fully managed Network File System (NFS) ideal for general file sharing. While its high-performance tiers offer solid throughput, it can hit scalability bottlenecks when thousands of clients access data concurrently in parallel.

Hyperdisk (Extreme Throughput): Next-generation block storage providing high IOPS and throughput, but it is primarily attached to a single VM with limited multi-VM sharing options.

Cloud Storage FUSE: Near-infinitely scalable object storage accessed like a file system via GCS FUSE. While it provides high throughput, metadata operations and latency may lag behind dedicated HPC file systems.

Storage Option

Primary Use Cases

Key Performance Metrics

Local SSD

Maximum performance temporary storage for individual VMs (volatile data)

Ultra-low latency,

Highest IOPS per VM

Managed Lustre

HPC, AI/ML training, media rendering, simulation, etc.

Workloads demanding extreme parallel and sequential I/O performance

Low latency,

Hundreds of GB/s throughput

Filestore

(High Scale)

Enterprise workloads, GKE persistent volumes, shared file systems

High availability,

High IOPS and solid throughput

(Lower compared to Lustre)

Persistent Disk

(Extreme/Hyperdisk)

Persistent block storage requiring high IOPS

Flexible performance tuning, persistence, reliability

Cloud Storage FUSE

AI/ML data loading (cost-effective), checkpoint storage,

Data lakes requiring virtually unlimited capacity

Virtually unlimited scalability, high sequential read throughput (though metadata operations and random I/O latency are higher than Lustre)

1-3. When Should You Choose Lustre?

With multiple high-performance storage options available, you might wonder: "When should I choose Lustre?" Managed Lustre is the ideal choice in the following scenarios:

Extreme Concurrency: Hundreds to thousands of clients (VMs, containers) must read or write to the same file system simultaneously.

Maximum Throughput Needs: You require tens to hundreds of GB/s of evenly distributed throughput that exceeds the limits of Filestore or individual disks.

Strict POSIX Compliance: You need to migrate legacy HPC applications directly to the cloud without modifying application code.

1.4 Use Case: Vertex AI Training Cluster and Managed Lustre

Vertex AI Training Cluster (VTC) allows you to provision and manage machine learning training clusters directly inside Google Cloud's fully managed Vertex AI platform. Because it is a fully managed service, it eliminates the operational overhead of requiring a dedicated internal AI infrastructure team to build and maintain large-scale HPC clusters. You can easily spin up and operate HPC clusters powered by AI accelerators such as GPUs and TPUs. For instance, you can run a cluster with thousands or tens of thousands of data center GPUs without shouldering infrastructure maintenance overhead and its associated costs.

Large-scale HPC clusters represent a prime use case for high-performance storage. In these environments, Managed Lustre is strongly recommended as the core storage option to maximize overall cluster efficiency:

Maximizing GPU/TPU Utilization (Eliminating I/O Bottlenecks): To prevent expensive GPUs and TPUs from sitting idle while waiting for data to load, Managed Lustre feeds training data to accelerators at ultra-high read throughput.

High-Speed Checkpointing: Training large language models (LLMs) requiring large-scale AI accelerators (e.g., GPUs) involves frequent checkpointing to save model states. Checkpointing creates massive burst write traffic; Managed Lustre's parallel write performance significantly cuts write times, boosting overall training throughput.

Configuration: When provisioning a VTC, selecting Managed Lustre as the storage option connects it directly to the cluster's internal VPC via Private Service Access, allowing you to use high-performance storage immediately without complex manual mount configurations.

For more information on VTC, which launched in Public Preview in Q4 2025, refer to the Vertex AI Training Cluster Overview; for details on VTC storage options, see Vertex AI > Storage.

Next: Building Managed Lustre Storage on Google Cloud (2/3)