File shares

Provisioning file shares for your instances

File share storage is currently in private preview and requires a quota to be added to your account. For access, contact your account team via Slack or email support@mlfoundry.com.

File shares provide convenient, high-performance file storage for your instances. File shares are automatically mounted to your instances as a directory under /mnt, and are readable and writable from multiple instances concurrently.

File shares are read-optimized. You are likely to see higher read speeds in practice.

Provisioning file share storage

To provision file share storage, go to Storage, click + Create storage and select File share for the type field.

  1. Select the region you would like to provision in.

  2. Select the size for your disk. Currently, each file share storage has a maximum size of 15TB.

When creating a file share, you must provide a unique Name which will be used for the folder where it will be automatically mounted inside your instance. For example, if you named your file share "training-model" it will be automatically mounted on the /mnt/training-model folder after your instance is launched.

Region
GPUs available

us-central1-a

NVIDIA H100

na-east1-a

NVIDIA A40

na-east1-b

NVIDIA A5000

eu-central1-a

NVIDIA A100

Attach storage to reservations and spot bids

Provisioned storage is available for selection while creating a reservation or spot bid. You can attach as many storage options as needed from the same region. If you select a region that does not have provisioned storage, it will not appear as an option.

Currently, it's not possible to attach storage to existing instances or shrink/expand disk size.

Accessing the disk from your instance

You can list all file shares attached to your instance with ls:

$ ls /mnt

You can read and write to any attached file share folders.

Performance & benchmarks

Actual performance for your file share is likely to vary greatly depending on your workload and instance configuration; however, below are representative benchmarks for a single 8x H100 instance:

Sequential Read BW
Read IOPS (4Kb)
Sequential Write BW
Write IOPS (4Kb)

12 Gbps

12,000

5 Gbps

4500

The numbers above are a snapshot meant for guidance. We are constantly making improvements to optimize performance. You can also run the benchmarks yourself from your instances using the following:

# sequential write bw
fio --name=my-job --group_reporting --time_based=1 --cpus_allowed_policy=split --runtime=10s --ramp_time=5s --size 20G --numjobs=32 --direct=1 --iodepth 8 --rw write --bs 1m
# sequential read bw
fio --name=my-job --group_reporting --time_based=1 --cpus_allowed_policy=split --runtime=10s --ramp_time=5s --size 20G --numjobs=32 --direct=1 --iodepth 8 --rw read --bs 1m
# write iops
fio --name=my-job --group_reporting --time_based=1 --cpus_allowed_policy=split --runtime=10s --ramp_time=5s --size 20G --numjobs=32 --direct=1 --iodepth 8 --rw write --bs 4k
# read iops
fio --name=my-job --group_reporting --time_based=1 --cpus_allowed_policy=split --runtime=10s --ramp_time=5s --size 20G --numjobs=32  --direct=1 --iodepth 8 --rw read --bs 4k

Quotas

Your account has limits on the amount of storage that can be created per region. Contact your account team via Slack or email support@mlfoundry.com to increase your quota. If you do not see the Storage tab in the main navigation area, please reach out to be

Last updated