A GPU cluster project for a training laboratory, covering the supply, implementation and tuning of a high-performance storage system, designed to give the client's AI applications stable, efficient and high-speed data storage support. UnicoNet tailored the solution to the client's performance requirements and delivered it as a platform that also meets specific requirements for job scheduling and resource management.

Client Background

The client is a training laboratory for AI and high-performance computing education. Unlike a cluster built for a single enterprise, this laboratory has to serve multiple class sessions, multiple projects and multiple users in parallel on one platform. It must guarantee compute and storage performance while providing clear resource allocation and usage boundaries.

Challenges

1. Concurrent compute and storage: The GPU cluster and back-end storage had to support highly concurrent reads and writes from many users and many tasks, while providing stable bandwidth for pre- and post-processing applications.
2. Job scheduling: Beyond performance, the platform had to implement combined policies for job queuing, priority escalation and preemption so that limited compute resources are used fairly and efficiently.
3. Private cloud management: Pre- and post-processing tasks had to run as cluster-native private cloud operations, giving users a single entry point to submit and retrieve jobs.

Solution

UnicoNet built a GPU cluster around 2124 all-flash storage and two 8-GPU servers, then deployed a job scheduling and resource management platform on top, forming a three-in-one architecture of compute, storage and scheduling:

• Compute tier: Two 8-GPU servers (8 × L20) provide compute power for AI training and inference.
• Storage tier: 2124 all-flash storage configured with 16 × 7.68 TB NVMe SSD, delivering low-latency, high-bandwidth data support for models and datasets.
• Scheduling tier: Job queue, priority, preemption and escalation policies together with user grouping and domain separation, delivering cluster-native private cloud operations.

Platform Interface

After delivery, administrators manage node status and resource allocation through a unified interface, while instructors and students submit pre- and post-processing and solver jobs through the private cloud entry point and view job status and resource usage in real time.

Figure 1: Pre- and post-processing application (solver) interface

Figure 2: GPU cluster node management interface (node name, IP address, BMC address and node status at a glance)

Figure 3: Job queue and resource allocation interface (job ID, status, partition, node and resource usage)

Configuration

Client Benefits

• GPU compute and all-flash storage are managed as one, so submitting, queuing and retrieving training tasks all happen on a single platform, markedly improving teaching efficiency.
• Job queuing, escalation and preemption policies are in place, so urgent and high-priority jobs obtain resources immediately instead of compute being held for long periods by low-efficiency tasks.
• Pre- and post-processing applications run as cluster-native private cloud operations, so users need not concern themselves with underlying node distribution, lowering the learning curve and operational cost.

Note: This case study is compiled from the client's project material. Interface screenshots illustrate platform functionality. Actual models, quantities and performance are subject to the final test report and quotation.