Data is sharded randomly across multiple SCM drives to increase throughput and decrease contention.
Data is written to two different SCM drives so no data is lost in the event of a SCM drive failure.
Metadata is updated in the internal data structure (tree) for consistency.
Comprehensive and Detailed 250 to 300 words of Explanation From Advanced Storage Solutions Architect documents and knowledge guide:
The write data path in HPE GreenLake for File Storage (powered by Alletra MP X10000 hardware and VAST Data software) follows a unique Disaggregated Shared-Everything (DASE) architecture. Unlike legacy NAS systems that use front-end caching or complex controller-to-controller talk, this solution leverages Storage Class Memory (SCM) as a persistent write buffer to provide high-sustained performance without the need for traditional data movement between tiers.
The process begins with sharding. When a NAS write request arrives, the system immediately shards the data randomly across multiple SCM drives in the cluster. This sharding is critical because it eliminates hot spots and contention by ensuring that no single drive or node becomes a bottleneck, effectively parallelizing the IO load across the entire storage fabric.
Once the sharding logic is determined, the data is physically written to the SCM tier. To ensure mission-critical resilience, every write is mirrored (written to two different SCM drives). Because SCM is non-volatile random-access memory (NVRAM), the write is persistent the moment it hits the media. This allows the system to send an immediate acknowledgement back to the client while protecting against a drive or node failure.
Finally, the metadata is updated in the internal data structure (the V-Tree). This step ensures the "View" of the file system remains consistent and that the global namespace reflects the newly written data. After this point, the data is asynchronously moved from SCM to high-capacity NVMe SSDs using wide-stripe erasure coding for long-term, efficient storage. This disaggregated flow allows the Alletra MP X10000 to scale performance and capacity independently while maintaining strict data integrity and consistency at AI-scale.
