A high-performance AI storage project 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 requirements for large-scale data processing and intelligent analysis, supporting its digital transformation and innovative applications.
Client Background
The client is a biotechnology company advancing AI-driven data processing and intelligent analysis. Such workloads are highly sensitive to throughput, access latency and capacity, and must remain stable while many nodes read and write simultaneously — a typical case of high-performance computing driving high-performance storage.
Project Challenges
- Highly concurrent access: The project environment comprises 60 AI terminals that read and write large volumes of data to the storage system at the same time, placing extreme demands on concurrent processing capability.
- High bandwidth: Phase 1 required storage write speed above 10 GB/s to match the data throughput of the AI applications.
- Capacity and scalability: Phase 1 capacity was 432 TB, and the architecture had to leave room for later expansion to avoid a second rebuild.
Solution
UnicoNet built a high-speed storage platform around 2124 all-flash storage, combined with a 100 Gb high-speed switching network and SAS expansion disks: the all-flash tier carries hot data and highly concurrent reads and writes, while the SAS disk tier provides large capacity. The two tiers are interconnected over dual-port 100 Gb links, and 60 front-end AI terminals connect to the storage network through switches to form a non-blocking data path.
Figure 1: AI storage platform topology (AI terminals → 100 Gb switch → hybrid storage → AI servers)
Measured Results
After implementation and tuning were completed, we ran concurrent read and write tests on the storage platform in the client's live environment. Throughput results for 8-thread concurrent write and 8-thread concurrent read are as follows:
| Test metric | 8-thread concurrent write (KB/sec) | 8-thread concurrent read (KB/sec) |
|---|---|---|
| Overall throughput (children view) | 2,459,058.75 | 5,521,198.75 |
| Overall throughput (parent view) | 14,263,490.72 | 5,515,548.87 |
| Minimum per-process throughput | 290,612.75 | 674,040.02 |
| Maximum per-process throughput | 315,437.75 | 699,506.04 |
| Average per-process throughput | 307,438.34 | 690,149.84 |
| Minimum transfer size (KB) | 3,864,576.00 | 4,042,752.00 |
Converting at 1 GB = 1024 MB, overall throughput in the parent view at 8-thread concurrent write was approximately 13.60 GB/s, meeting and exceeding the Phase 1 requirement of above 10 GB/s. At 8-thread concurrent read, overall throughput was approximately 5.27 GB/s, with stable read and write behaviour.
Figure 2: Raw output of the concurrent read and write test (8-thread concurrent write above, 8-thread concurrent read below)
Configuration
| Item | Configuration |
|---|---|
| Storage system | 2124 all-flash storage |
| Flash drives | 8 × 7.68 TB NVMe SSD |
| Expansion disks | 18 TB × 24 SAS disks |
| Network interface | Dual-port 100 Gb |
| Access scale | 60 AI terminals with concurrent read and write |
Client Benefits
- The write requirement of above 10 GB/s was met in Phase 1, so concurrent access from 60 AI terminals is no longer a bottleneck in the data processing pipeline.
- The 432 TB Phase 1 capacity combined with an expandable architecture means later expansion needs no redesign, protecting the existing investment.
- From solution design and equipment supply to commissioning and tuning, delivery came from a single vendor with clear accountability, shortening the project timeline.
Note: The measured figures above are concurrent read and write results from a specific test environment. Actual performance varies with application type, file size, node count and network conditions, and is provided for reference only. Final configuration and delivery schedules are subject to the actual test report and quotation.