Dell unveils AI and analytics storage appliance

Dell has produced the ObjectScale XF960, an all-flash, scale-out appliance for generative AI and real-time analytics work.

Update: Removed VSAN and Tanzu support from ObjectScale. Added updated XF960 throughput numbers; 9 November 2023.

ObjectScale is Dell’s software-defined object storage system. Now it is available as a turnkey hardware appliance with limited scalability. The XF960 is the first member of an ObjectScale X-Series family and based on a PowerEdge server running ObjectScale v1.3 software. The system includes servers, network switches, rack-mount equipment, and appropriate power cables.

According to an XF960 spec sheet, each node features dual Intel Xeon CPUs within a 2RU chassis and can support up to 737.3 TB of raw capacity. A cluster has a minimum of four nodes, five nodes for high-availability, and a maximum of 16, with up to 11,796 TB of raw capacity (11.8 PB). Its features include:

  • Dual Intel gen 4 Xeon SP Processors with 32 cores per processor
  • 256 GB of memory
  • 24 x NVMe TLC 30.72 TB drives (for the initial release)
  • Dual mirrored boot drives for redundancy
  • Network connectivity with dual 100 GbE back-end and dual 25 GbE front-end speeds.

The ObjectScale software is a distributed, microservices-based, multi-node, scale-out, and multi-tenant object storage system with a single global namespace supporting the S3 API.  It has the same code base as Dell’s ECS software. More than 1,000 nodes are supported, from a three-node start, with no specific architectural limit. The XF960’s scale-out limits are a function of the hardware and not the ObjectScale software.

ObjectScale runs on customer-managed SUSE Linux Enterprise Server (SLES) operating system and Red Hat Open Shift. The v1.3 release adds new S3 bucket logging and enhanced S3 replication. Replicated objects can now be as large as 50TiB, with support for up to 50,000 parts. v1.3 supports multipart upload (MPU) using the S3 protocol. MPU enables the construction of large objects from smaller parts that can be uploaded at any time and in any order.

v1.3 enables Object Lock to be applied at the bucket level, in the absence of versioning. When enabled without versioning, Object Lock applies at the bucket level instead of to individual versions within the bucket. Consequently, overwrite and delete operations are effectively prevented. A v1.3 ObjectScale blog offers more detailed information.

A Dell XF960 blog says it’s a fully integrated, turnkey system and “the world’s most powerful object storage appliance custom-built on Kubernetes.” It envisages software-defined ObjectScale systems at edge sites sending data to XF960 appliances in the datacenter.

We’re told customers can combine Starburst, the Trino-based data lake offering, with ObjectScale “to set the foundation for a high-performance, infinitely scalable lakehouse on-premises and across multi-cloud environments.” However, the XF960 is not infinitely scalable as it is limited to 16 nodes for the initial release, due to additional networking qualifications needed for larger scale-out.

What we have here are Dell object storage nodes that can deliver can deliver up to 5.7 GBps reads and 5.0 GBps writes on http per node. This is an object store used for primary, tier 1, S3 data storage feeding data direct to AI/ML and analytics processes.

Object storage supplier Cloudian can deliver 17.7 GBps writes and 25 GBps reads from a six node all-flash cluster. That’s equivalent to 2.95 GBps read and 4.2 GBps read speeds, slower than the XF960.

MinIO promotes object storage as a primary storage tier but we don’t have any equivalent per-node read and write throughput speed numbers for comparison with Dell and Cloudian.

A Dell blog tells us what Dell thinks about the XF960.