Docs

Elastic Compute Service

Elastic Block Storage

ML disks

An ML disk is a read-only volume that does not allow writing. It can be attached to multiple ECS instances simultaneously. This topic covers the use cases, benefits, specifications, and creation process of ML disks.

tip

ML disks are currently in closed beta. Contact your account manager to have your account added to the allowlist and start using it.

Use cases

ML disks have high throughput and multi-attach capabilities, making them ideal for the following scenarios:

AI and machine learning: They can be used to store models in the inference phase, accelerating model loading and significantly reducing service latency and startup time.

Container image distribution: They store container images and support concurrent reads from multiple instances. This accelerates image distribution and shortens container and application startup times.

Benefits

Benefit	Description
Flexibility	ML disks have a minimum capacity of 10 GiB and allow you to quickly create and delete resources for flexible management.
Ultra-high performance	A single ML disk can support up to 100 ECS instances or container pods simultaneously, delivering GB/s-level aggregated throughput. This feature meets the demands of high-throughput scenarios, such as model training and image distribution.
Cost optimization	Data is stored once and is accessible globally — you only pay for storage once. You are billed for throughput performance on a pay-as-you-go basis, so there is no need for upfront provisioning.

Volume performance

Attribute	Performance
Maximum read throughput per volume (MB/s)	Maximum aggregated read throughput: 100 GB/s. Maximum read throughput per attachment: 1 GB/s.
Average latency of single-path random writes (block size = 4K)	Sub-millisecond.
Maximum capacity per volume (GiB)	The capacity depends on the size of the source snapshot. System volume: Not applicable. Data volume: 10–65,536.
Maximum number of instances to which a volume can be attached	100.
API name	ML_Disk

Notes

Category	Notes
Instance type	You can attach ML disks to the following instance types: Heterogeneous Computing: hpcpni3ln.
Attachment	You can attach an ML disk to up to 50 instances at a time.
Unsupported operations	ML disks do not support the following operations: Extend a volume. Roll back volumes using snapshots. Encrypt a volume. Change the volume type. All snapshot operations except import. Reinitialize a volume. Attach a volume when creating an instance.
Constraints	ML disks only support read-only access. Writes are not allowed.

Creating an ML disk

Console

API

In the ECS console, go to the Volumes page and click "Create volume".

On the "Create volume" page, under "Storage Specifications > Storage", set the data volume type to "ML_Disk".

You can enable "Setting performance limit" to limit throughput performance based on your needs. The valid range is 1–102,400 MB/s.

Click "Create from snapshot" and specify the "Data source". You can only create an ML disk from a snapshot. Creating an empty ML disk is not supported.

Follow the on-screen instructions to configure the remaining parameters. For more information, see Creating a volume.

Billing notes

ML disks are billed based on capacity and throughput. Pay-as-you-go billing is the only option. Subscription billing is not available. The billing details are as follows:

Billable item	Formula	Billing method	Unit price	Description
Volume capacity	Cost = Unit price × Provisioned storage capacity × Usage time	Pay-as-you-go	¥0.0021/GiB/hour	You are charged for capacity usage once you create an ML disk. For example, using a 100 GiB ML disk for two hours incurs a capacity cost of ¥0.42 (0.0021 × 100 × 2).
Throughput	Throughput cost for a given hour = Average throughput per second within the hour (MB/s) × Unit price	Pay-as-you-go	¥0.056/MBps/hour	The cost of throughput is based on your actual usage. For example, if the average throughput per second for a given hour is 50 MB/s, the cost for that hour would be ¥2.8 (1 × 50 × 0.056).

Last updated: 2026.07.17 14:31:51