Boston Flash-IO Talyn

Editorial Type: Review Date: 09-2019 Views: 1,362 Tags: Storage, NVMe, SSD, Boston, Flash-IO Talyn, NVMesh
Enterprises are rapidly seeing the benefits of ultra-fast NVMe storage for demanding applications such as IoT, Big Data analytics and artificial intelligent (AI). However, traditional storage solutions fall apart when sharing these resources across multiple remote clients as latency and performance inevitably tumble

Boston's Flash-IO Talyn overcomes these issues by presenting centralised, shared NVMe storage with near local latency and speeds. The power behind the Talyn's throne comes courtesy of Excelero's innovative NVMesh software which Boston has deployed on a range of 1U and 2U Supermicro server-grade platforms equipped with Mellanox Connect-X 5 100Gbits/sec Infiniband and Ethernet adapters.

NVMesh runs on Linux and the Talyn is ready to go out of the box with Excelero's target module and TOMA (Topology Manager) preinstalled. Multiple appliances can be placed in clusters where the TOMA forms a quorum, elects one node to be the leader and distributes topology changes across all cluster members.

For more versatility, you can have up to four clusters within one appliance. In all cases, NVMesh employs Excelero's Elastic NVMe technology to provision a distributed block level architecture where NVMe RAID arrays are spread across nodes for enhanced scalability and redundancy.

Remote systems only require Excelero's Linux client block driver installed which uses insignificant amounts of local memory as data operations are synchronous and do not utilise a cache. Connection choices are good as the driver supports 100Gbits/sec Mellanox InfiniBand and Ethernet NICs.

Remote management services run on a Linux platform and for high availability, multiple hosts can be placed in their own cluster. This provides full management redundancy and hourly snapshots are also performed for swift configuration restores.

The web console presents a smart dashboard showing cluster members, NVMe SSDs and connected clients plus an allocation chart with details on redundancy levels, virtual volumes and available space. The console is well designed and we were impressed with the transparency of the underlying NVMesh technologies.

Selected NVMe SSDs are placed in provisioning groups where you choose a RAID array, create virtual volumes and assign them to clients. NVMesh requires a minimum of two nodes and blocks in the group are distributed across them so if an entire node fails, access to storage is maintained.

Client configuration couldn't be easier as a single file tells them where the management service is. They automatically attach to assigned virtual volumes which appear as standard block devices that can be mounted and configured with file systems.

Client access to storage is achieved using Excelero's patented RDDA (remote direct drive access) protocol which sets up pathways on RDMA-capable NICs. As IO is sent directly to and from the client, no extra CPU cycles are consumed on either side with Excelero claiming this only adds 5 microseconds to the response time for round-trip operations.

Perfect for serving up low-latency storage to GPU artificial intelligence (AI) and machine learning (ML) farms, the Talyn delivers an impressive performance. Using an NVIDIA DGX-1 deep-learning and analytics platform, Boston demonstrated an impressive throughput of 67GB/s for sequential reads and 1.2 million IOPs for random reads.

You can see all the action from the web console statistics page with graphs showing overall throughput and IOPS along with the most active clients. Roles can be used to provide admin and observer access to the console while Excelero's RESTful API allows third party products such as Ansible to automate storage configuration.

Product: Flash-IO Talyn
Supplier: Boston
Web site: www.boston.co.uk
Sales: +44 (0)1727 876 100

Verdict: Boston's Flash-IO Talyn takes software-defined storage (SDS) to the next level by provisioning scale-out block storage with in-server Flash performance. It's versatile, scalable and remarkably easy to deploy as it doesn't require applications and workflows to be modified to access it.