12 GB/s (read) per Petabyte on HDD: Cloudian HyperStore Proves Capacity and Performance Can Coexist

Unstructured data keeps growing, and most of it lands on hard drives. HDD media still offers the lowest cost per terabyte for petabyte-scale storage. But HDDs carry a reputation: plenty of capacity, not much speed. As drives grow past 20 TB each, that worry gets louder. Bigger spindles are not faster spindles, and many teams assume they will need an all-flash tier to keep their applications fed.

There’s a better way. In a recent customer acceptance benchmark, Cloudian HyperStore® delivered 12 GB/s of read throughput and 3.8 GB/s of write throughput for every petabyte of usable capacity, on a 24-node, all-HDD cluster. Capacity and performance came from the same drives.

Cloudian HyperStore at a Glance

Cloudian HyperStore is S3-compatible object storage for on-premises and hybrid cloud environments. It gives enterprises the simplicity of the public cloud’s storage API with the control, security and cost profile of their own data centers. Its modular, distributed architecture scales non-disruptively from terabytes to exabytes, one node at a time.

A few traits define the platform:

• Shared-nothing, peer-to-peer design. Every node stores data and serves S3 requests. There is no separate gateway or metadata tier to become a bottleneck.
• Flexible data protection. Replication and erasure coding are set per storage policy, so each workload gets the right balance of durability, capacity and performance.
• Built-in load balancing. HyperBalance™ Fusion spreads client traffic across nodes as part of the cluster, with no external appliance to buy or manage.
• Proven for AI. HyperStore has achieved NVIDIA-Certified Storage designation and supports data lakes, analytics, backup and media workloads alongside AI pipelines.

Cloudian’s Test: Real Hardware, Real Workloads

We ran the benchmark on a customer’s production cluster as part of their acceptance testing. The goal was simple: measure how much S3 throughput each petabyte of usable HDD capacity could deliver under heavy, concurrent load.

To keep the test honest, read datasets were sized well beyond the cluster’s memory, and caches were cleared before every read run. Every result you see below came from disk, not RAM.

Results: 3.8 GB/s Write and 12 GB/s Read per Usable Petabyte

At peak, the cluster sustained 63.6 GB/s of reads and 20.2 GB/s of writes over HTTPS with 10 MiB objects. Spread across 5.28 PB of usable capacity, that is 12.0 GB/s of reads and 3.8 GB/s of writes per usable petabyte.

Cloudian customer acceptance benchmark, October 2026 · RF3, HTTPS · usable = raw HDD capacity / 3

Reads stayed above 10 GB/s per usable petabyte at every object size we tested. Writes cleared 3 GB/s per usable petabyte for every object of 8 MiB and larger. Smaller objects write more slowly per byte, since each request carries fixed per-object work, and landed at 1.9 to 2.4 GB/s per usable petabyte at 1 to 2 MiB.

Security didn’t cost the headline either. At 10 MiB, HTTPS and server-side encryption both held above 3 GB/s per usable petabyte for writes and 10 GB/s for reads.

Why HyperStore Keeps Every Spindle Busy

HDDs are not slow. At peak, each of the cluster’s 720 drives delivered about 88 MB/s of reads, and writes put about 84 MB/s per drive onto disk once replicas are counted. For high-capacity drives under heavy, concurrent S3 load, that is close to what the media can sustain.

Three design choices make that possible:

• Every node serves S3 data. HyperStore’s peer-to-peer architecture means all 24 nodes accept requests. HyperBalance Fusion spread 4,800 concurrent requests evenly across the cluster, and per-node traffic matched total throughput within a few percent.
• Data spreads across every drive. Objects and their replicas are distributed across nodes and disks, so a single large stream becomes parallel I/O across the cluster instead of a hot spot on a few drives.
• Balanced, dense nodes. Each node pairs 30 drives with 96 cores and 512 GB of RAM. That is enough compute to handle TLS and encryption without starving the disks.

RF3 replication adds its own advantage. Writes need no erasure coding math, and every object exists as three full copies, which gives the system more places to serve each read from.

What This Means for Your Storage Plan

When HDD capacity delivers 3 GB/s or more per usable petabyte, the capacity you buy for retention also buys real bandwidth. Many backup, analytics, media and AI data lake workloads can be sized on capacity first. An all-flash tier is often not required to achieve a throughput target.

• Performance scales with capacity. In HyperStore’s shared-nothing design, every HDD node adds both petabytes and bandwidth. Grow the cluster and throughput grows with it.
• RF3 replication suits hot, write-heavy data. It trades raw capacity for bandwidth and simplicity, a good fit for backup landing zones, media ingest and AI training sets. Budget raw capacity at three times usable.
• Large objects get the most out of HDD. From 8 MiB up, writes held at 3.5 GB/s per usable petabyte or more. Workloads that write large objects or multipart uploads get the full benefit.
• Add flash only where it pays. Objects under 1 MiB are bound by drive seek time. HyperStore storage policies can place them on a small flash tier while large objects stay on HDD.

The Bottom Line

HDD media is not the performance bottleneck many assume it is. On standard high-capacity hardware, Cloudian HyperStore turned 720 spindles into 63.6 GB/s of reads and 20.2 GB/s of writes over HTTPS: 12 GB/s and 3.8 GB/s for every usable petabyte. With HyperStore, you no longer have to choose between capacity and performance.

Planning a multi-petabyte object store? Talk with a Cloudian Solutions Expert to see how these results map to your workload, or request a proof of concept on your own data.

Test notes: Throughput is measured S3 data throughput at peak concurrency, averaged over two repeats and converted from MiB/s to GB/s. Usable capacity is raw HDD capacity (24 nodes x 30 x 22 TB = 15.84 PB) divided by three for RF3, before free space reserves.