Pure Storage: New highly scalable high-performance storage FlashBlade//EXA

With the FlashBlade//EXA, Pure Storage has introduced an all-in-one storage system for AI and HPC applications that is designed to be almost infinitely scalable in all directions. The manufacturer clusters any number of controller nodes into a metadata core, which communicates with the HPC or AI cluster via NFSv4.1 over TCP. The metadata core is therefore designed for high parallelism and enormous quantities of metadata operations.

It is connected to the data nodes via a control panel, each of which holds 24 DFMs (Direct Flash Modules) with 75 or 150 TByte flash, i.e., 1.8 or 3.6 PByte. Furthermore, almost unlimited commercially available data nodes from third-party providers can also be connected and managed with the in-house Purity operating system.

Read performance of more than 10 TByte/s?

Once the metadata core has given the requesting compute node the address of the data, it can access it directly via NFSv3 or RDMA. In initial scenarios, Pure Storage promises a read performance of more than 10 TByte/s in a single namespace. However, this would require an expansion level of three full racks, as Pure specifies a performance density of 3.4 TByte per rack. Nvidia's high-performance hardware such as the Mellanox ConnectX NICs, Spectrum switches and LinkX cables can also be integrated into the cluster, while Accelerated Communications Libraries accelerate communication between the nodes.

With their massively parallel processing, AI and HPC applications place different demands on storage systems and clusters than classic enterprise applications. These include parallel read and write accesses, high metadata performance, extremely low latencies, asynchronous check pointing and high throughput. With the FlashBlade//EXA, Pure Storage also wants to serve AI training and AI inference applications in addition to classic HPC applications.

Read also

Pure Storage: New systems to herald the end of hard disks

(sun)