Skip to main content
Local storage on CoreWeave Nodes uses a software RAID array assembled from the Node’s local NVMe drives, mounted at /mnt/local. The RAID layout is managed by CoreWeave and is not customer-configurable. This article explains the RAID configuration, why RAID0 is not available, and what to expect for performance.

​
RAID layout by drive count

The RAID layout depends on how many local NVMe drives the Node has, which varies by instance type:
  • Nodes with four or more drives use RAID10 (a striped mirror). Most GPU Nodes have eight drives in an 8-drive RAID10 array.
  • Nodes with two or three drives use RAID1 (a mirror).
  • Some instance types provide a single drive, presented as one device with no RAID.
RAID10 combines mirroring (RAID1) with striping (RAID0): each write is mirrored across drives, and reads and writes are striped across the array for throughput. Local storage is mounted at /mnt/local inside your containers. For more details on the layout, throughput, and benchmarking, see Local storage.

​
Why RAID0 is not available

RAID0 is not available because of data integrity constraints. RAID0 stripes data across drives without redundancy, so a single drive failure results in total data loss for the array. RAID1 and RAID10 provide fault tolerance by mirroring data, allowing the array to survive individual drive failures without data loss. Because CoreWeave manages the underlying hardware and multiple workloads may share the same physical infrastructure, a mirrored layout is enforced to protect against hardware failures.

​
Expected performance characteristics

Throughput depends on the RAID layout, and therefore on the instance type:
  • A striped RAID10 array, used by most GPU Nodes, delivers high parallel read and write throughput that scales with the number of drives and the concurrency of the workload.
  • A RAID1 mirror, or a single drive, delivers lower throughput. Reads can split across the two drives of a mirror, but writes are about the speed of a single drive because they are mirrored rather than striped.
To measure throughput accurately, benchmark with a concurrent tool such as fio with multiple jobs. A single-stream tool such as dd is latency-bound and understates a striped array. For write-intensive workloads, also use large sequential writes where possible, and avoid many small random writes, which amplify the mirroring overhead.

​
When to contact support

Contact CoreWeave support if:
  • You need guidance on optimizing write performance for your workload on local storage.
  • You are experiencing lower-than-expected local storage throughput.
  • You have questions about local storage availability on a specific instance type.
Administrator
Last modified on June 26, 2026