Skip to main content

High-Performance Storage (HPS)

An HPS device is a persistent, NFS-like filesystem you provision, size, and attach to jobs and workstations. It provides a central shared file store with SSH and SFTP access, reachable from any cluster for compute, post-processing, or visualization.

Your devices are listed under Storage, with the space each one has left and the region it lives in.

The Space column carries the available figure underneath the provisioned size, which is the number to watch.

HPS or CFS?​

Both mount a shared filesystem into your jobs, so the question comes up often. They suit different situations and neither replaces the other.

Choose HPS when:

  • You need to reach the storage directly. An HPS answers SSH and SFTP, so you can rsync into it or open it in VS Code Remote. A CFS does neither.
  • You want several independent devices. You can run as many as you like, each sized, shared, and terminated separately, where a workspace gets one CFS at most.
  • Your workstations run Windows. CFS mounts on Linux only.
  • The rest of the workspace must stay unconstrained. An HPS lives in one region and a job that attaches it has to run there too, but only that job. A CFS is the opposite: it pins every job in the workspace to its region, mounted or not.
  • You want to control the thing yourself. You set the size, the walltime, and the alert thresholds, and you can extend it later without involving Rescale.

Choose CFS when:

  • The I/O is the bottleneck. Lustre substantially outperforms NFS on parallel I/O, and HPS is built for low and moderate bandwidth rather than heavy throughput.
  • The data needs to outlive any schedule. A CFS has no walltime and runs no compute instance, so nothing terminates it and nothing accrues hourly.
  • Many jobs read it at once. An HPS serves files over NFS and that imposes node limits you can hit without warning.
  • A whole team shares it. CFS shares through Rescale Projects across the workspace, with no per-device setup and no per-device sharing step.

HPS versus CFS versus Rescale Files​

Side by side, the three differ most in who provisions them, how they bill, and whether they end on their own:

Rescale FilesCFSHPS
TypeObject storageLustre filesystemNFS-like filesystem
ScopePer userWorkspace-wide, one per workspacePer device, many allowed
Constrains where jobs runNoEvery job in the workspaceThe jobs that attach it
You provision itNoRescale doesYes
BillingBy GiB storedBy provisioned capacityHourly while active
Has a walltimeNoNoYes
SSH / SFTP accessNoNoYes
Windows workstationsn/aNoYes
Indefinite retentionYesYesNo — 365-day cap

The lifecycle to plan around​

HPS devices are billed hourly and terminate on walltime, which makes them unlike the other two storage options.

StageBehavior
ProvisioningUp to 15 minutes before the device is usable
ActiveBilled hourly — see what it costs
Walltime reachedFiles in work/ upload to Rescale Files, then the device terminates
Budget exhaustedA strict project budget running out terminates the device
A device can terminate without you asking

Walltime and an exhausted project budget each kill an HPS device. Walltime is capped at 365 days and counts from creation, so once a device has started it cannot be pushed beyond 365 days from its cluster's start time. A device that started nine months ago accepts a new walltime of under three months.

You can still change the walltime after launch, within that bound, from the device's Settings tab. Rescale also emails you before a device syncs and stops, if the walltime alert is on, which it is by default on new devices.

HPS is not permanent storage

The 365-day cap is deliberate. HPS exists so jobs don't have to re-copy data, not to hold data indefinitely. For long-term retention use Rescale Files.

What the shutdown sync does and doesn't preserve​

Only files under the cluster's work/ directory are synced to Rescale Files at shutdown. And the sync is lossy in a way that surprises people:

Directory structure and permissions are lost

Files synced at shutdown land in Rescale Files as input files with only their name and size preserved. Relative paths, timestamps, and POSIX permissions are not kept.

A carefully organized 50,000-file tree comes back as a flat list. If structure matters, archive it yourself, tar it up before the device terminates.

Job output is unaffected: files in a job's work/ upload on job completion exactly as they would without an HPS.

What an HPS actually costs​

Unlike the other storage options, an HPS runs a compute instance, so several charges accrue at once:

ComponentBasis
ComputeThe coretype's on-demand rate, for as long as the device is active
Platform licenseCharged on the HPS compute, for Pricing 3.0 customers
StoragePer TB per hour
IP addressOn AWS devices

The setup form totals these for you as you size the device, and the same summary stays on the device's Settings tab afterwards.

The summary updates as you change the size or coretype, so you can see the hourly cost before you submit.
This is why an idle HPS is expensive

A storage device is billed like a running job, not like object storage. An HPS you've forgotten about accrues compute charges around the clock.

Extending capacity adds to the storage charge for the remaining lifetime. If a device is only occasionally needed, terminating it and letting the shutdown sync move data to Rescale Files is usually cheaper than leaving it up.

What the storage underneath is​

An HPS presents NFS to your jobs, over block storage from the cloud provider: gp3 EBS volumes on AWS and Premium SSD LRS on Azure. Rescale arranges those disks as RAID 0, the same as every other Rescale cluster disk array.

RAID 0 is worth knowing about, because it stripes for speed and keeps no redundancy. Combined with the 365-day cap, it is the reason an HPS is a working store rather than a place to keep the only copy of anything. Throughput and IOPS both scale with disk size on AWS, so a larger device is also a faster one.

Pages in this section​