HPE and Cloudian have joined forces to offer a solution that can save 90% on object storage capacity, while still delivering the superior S3 API fidelity of the world’s leading object storage platform.
Modern data platforms have converged on S3-compatible object storage as their preferred foundation. The reasons are clear: horizontal scalability, open API compatibility, and operational simplicity. At the same time, enterprises increasingly want to maximize physical storage efficiency — taking advantage of the inline deduplication and compression capabilities available in today’s NVMe storage arrays.
Traditionally, these have been separate conversations. Object storage platforms are typically deployed on direct-attached disks, while deduplication and compression have lived in purpose-built primary storage arrays. The question this joint Cloudian and HPE validation set out to answer is straightforward: what happens when you combine the best of both?
|
The opportunity: By running Cloudian HyperStore S3-compatible object storage on top of an HPE Alletra NVMe array with inline deduplication and compression, organizations can capture substantial capacity savings, without changing a single S3 API call or application configuration. |
The Architecture: Cloudian HyperStore on HPE Alletra NVMe
The joint Cloudian and HPE solution layers two complementary technologies: Cloudian HyperStore provides the S3 abstraction layer — scale-out, policy-driven, fully API-compatible — while HPE Alletra Storage MP provides the physical storage substrate with inline data reduction beneath it. Each layer is independently optimized.
| Cloudian HyperStore | HPE Alletra Storage MP B10240 |
| — Full S3 API compatibility
— Scale-out, software-defined architecture — Policy-based lifecycle and tiering — Multi-site replication (sync/async) — Kubernetes and AI platform integration — Runs on HPE ProLiant DL380 Gen11 |
— All-NVMe array, sub-millisecond latency
— Inline deduplication + compression — Six-nines (99.9999%) availability — 64 Gb Fibre Channel fabric connectivity — Independent capacity scaling — HPE InfoSight predictive analytics |
HyperStore operates transparently on top of the HPE Alletra SAN backend. From the S3 layer’s perspective, it is simply writing to block storage — all S3 APIs, replication policies, lifecycle rules, and multi-site capabilities are fully preserved. Beneath the surface, Alletra applies inline deduplication and compression before every write, consuming physical capacity only for the unique, compressed data blocks.
Crucially, the two layers scale independently: HyperStore scales S3 throughput by adding compute nodes, while HPE Alletra MP scales capacity and storage performance by adding NVMe enclosures — without the over-provisioning that plagues monolithic storage architectures.
Reference Architecture
Figure 1 — Cloudian HyperStore + HPE Alletra B10000 Reference Architecture: 3-node HyperStore cluster on HPE ProLiant DL380 Gen11, connected to HPE Alletra B10240 via 64 Gb Fibre Channel. Client traffic distributed via HAProxy across all nodes.
Performance Test Results
All performance testing was conducted using Intel Gosbench, a widely adopted object storage benchmark that generates high-concurrency, realistic S3 workloads. Gosbench was executed simultaneously on all three HyperStore nodes, with HAProxy distributing requests to prevent local affinity and ensure accurate cluster-wide measurements.
Read Throughput: Linear Scaling to Cluster Saturation
Figure 2 — HyperStore Read Bandwidth (Gosbench, MB/s vs. Worker Threads). Throughput scales linearly with worker threads until cluster-wide saturation, confirming no configuration bottlenecks.
Read throughput demonstrated strong and consistent linear scaling with worker thread count. The performance plateau at saturation reflects a fully optimized cluster — the ceiling is the available compute and network bandwidth, not a software or storage bottleneck.
Write Throughput: Consistent Under Replication Overhead
Figure 3 — HyperStore Write Bandwidth (Gosbench, MB/s vs. Worker Threads). Consistent write throughput maintained even with RF3 replication enabled across all three nodes.
Write performance remained stable and predictable even with Cloudian’s replication factor of 3 enabled across the cluster. The HPE Alletra NVMe backend absorbed write-intensive bursts without throttling, allowing HyperStore to sustain strong write bandwidth under load.
IOPS: Small-Object Concurrency for AI Inference Workloads
Figure 4 — HyperStore IOPS Results (Gosbench, IOPS vs. Worker Threads). Strong transaction scalability under increasing concurrency, validating suitability for metadata-intensive AI inference pipelines.
IOPS testing validated the architecture’s suitability for small-object, high-transaction workloads — characteristic of AI inference pipelines and metadata-heavy operations. Latency remained in the sub-millisecond to low-millisecond range across moderate concurrency levels, with predictable linear increase as the cluster approached saturation.
Data Reduction Results: The Headline Finding
Performance numbers are table stakes for enterprise storage. The finding that differentiates this architecture is what happened on the back end — the data reduction efficiency delivered by HPE Alletra’s inline deduplication and compression beneath a live HyperStore object storage workload.
Test 1 — Mixed Real-World Dataset: 22:1 Reduction
The first test used a representative enterprise dataset: operating system ISO images, virtualization packages, video media files, and application binaries — content typical of backup, archive, and media repository workloads. These file types share significant repeated data blocks both within and across objects, making them strong candidates for block-level deduplication.
Figure 5 — HPE Alletra Capacity Analytics: Test 1 Results. 113.9 TiB of logical data reduced to 10.9 TiB physical — a 22:1 data reduction ratio representing over 90% storage savings.
| Test 1 result: 113.9 TiB of logical object data → 10.9 TiB of physical storage consumed. Data reduction ratio: 22:1. Storage savings: >90%. |
Test 2 — Gosbench Synthetic Dataset: >25:1 Reduction
The second test used programmatically generated Gosbench synthetic objects — millions of objects across mixed sizes, written under high-concurrency conditions. These datasets, representative of AI training and analytics workloads, contain highly compressible and deduplicated content, establishing an upper bound on data reduction efficiency for this class of workload.
Figure 6 — HPE Alletra Data Reduction Results: Test 2. 48.6 TiB of logical data reduced to 4.0 TiB physical — exceeding 25:1 data reduction, representing over 96% storage savings.
| Test 2 result: 48.6 TiB of logical object data → 4.0 TiB of physical storage consumed. Data reduction ratio: >25:1. Storage savings: >96%. |
Summary: Both Tests Side by Side
| Metric | Test 1: Mixed Real-World | Test 2: Gosbench Synthetic |
| Logical Data Written | 113.9 TiB | 48.6 TiB |
| Physical Storage Used | 10.9 TiB | 4.0 TiB |
| Data Reduction Ratio | 22:1 | >25:1 |
| Storage Savings | >90% | >96% |
| S3 Compatibility Impact | None — fully transparent | None — fully transparent |
| HyperStore Config Changes | None required | None required |
A critical architectural note: inline deduplication operates at the block level beneath HyperStore, transparently reducing physical capacity across stored objects. No changes to HyperStore configuration or S3 operations are required.
Summary: What This Architecture Actually Delivers
The validated test results confirm that it is now possible to run enterprise-class S3-compatible object storage — with full API compatibility, scale-out architecture, and multi-site replication — while simultaneously achieving data reduction ratios that have historically been the exclusive domain of primary storage arrays.
| Key takeaway:
For organizations managing hundreds of terabytes or petabytes of object data, a 22:1 to 25:1 reduction ratio translates directly into fewer drives, fewer enclosures, lower power consumption, and a dramatically reduced data center footprint — with no application changes and no S3 capability trade-offs. |
The architecture is purpose-built for the workloads that are driving storage growth today:
| 🤖 AI Data Lakes
High-bandwidth object storage for training datasets and inference artifacts, with data reduction maximizing physical efficiency. |
💾 Backup & Archive
Common VM and file backup sets achieve 20:1+ deduplication ratios, dramatically reducing long-term retention costs. |
| 🎬 Media Repositories
Large video and image asset libraries managed via S3 compatibility, with direct integration into media workflows and CDN pipelines. |
☁️ Cloud-Native Applications
On-premises S3 with full data sovereignty — no API changes, seamless integration with AWS, Azure, and GCP. |
Organizations with existing HPE Alletra MP deployments can layer HyperStore directly on available capacity — enabling object storage workloads on infrastructure already in place, without a new storage procurement cycle.
Download the full technical white paper for complete hardware specifications, detailed benchmark methodology, and extended performance and capacity analytics from both test datasets.
cloudian.com | hpe.com
