This article is more than 1 year old
Excelero stumbles, squinting, into sunlight clutching its remote direct-access NVMe kit
Drink data from networked boxes with no CPU overhead
Storage startup Excelero slides out of stealth today with its NVMesh v1.1 Server SAN software, which can drink data directly from a remote node, bypassing that node's CPU.
Excelero is active in the NVMe over Fabrics area, in which remote direct access to a storage resource is almost as fast as direct local access, and has, in advance of the NVMe over Fabrics protocol, developed its own remote direct drive access (RDDA) protocol.
NVMesh is block storage access software that we envisage running in host server nodes fitted with NVMe flash drives. These nodes collectively provide a SAN resource and can underpin a hyperconverged product offering with much lower cross-node storage access resource use.
When one node needs to access data on another's NVMe storage, the NVMesh software makes a request to the remote node's NIC, using the RDDA protocol, and data is read or written remotely with no remote node CPU involvement.
This means remote node CPU cycles are not taken away from running application code to handle a storage access. As you can imagine, this builds on the inherent quickness of NVMe flash drive access and an NVMesh-equipped hyperconverged system should out-perform any other hyper-converged system and scale to larger node counts because of it. The company claims that performance scales almost linearly as nodes are added.
Excelereo performance scalability example
It also says that drive utilization can head toward 90 per cent – well in advance, it says, of today's 25 per cent or so utilization.
As remote data access is so fast, with roughly 5us being added for remote node access, data locality becomes a much less relevant issue. In an NVMesh hyperconverged infrastructure you could say all data is local throughout the system.
The software is structured with:
- Management layer handling provisioning and monitoring.
- Storage services such as data protection, multi-pathing and logical volumes.
- Hardware abstraction to create a pool of block storage out of non-volatile devices and networking.
The hardware abstraction layer supports RDDA and NVMeF for transport.
Excelero sees Web and cloud, innovative enterprises, HPC, and Media and Entertainment as its broad-brush market areas. Example apps include analytics and realtime analytics, databases, containers, visualization, burst buffer, genomics, simulation and electronic design automation.
Listed partners include Micron, Intel, Broadcom and Mellanox.
Excelero is talking to one prospective hyperscale customer about a system scaling to 100,000 nodes. Existing customers include NASA Ames, PayPal, Predix (GE's IoT business) for on-premises IoT data handling, and Hulu.
Predix had apparently already decided on a server-based flash route for its on-premises IoT data wrangling kit and switched to Excelero. One thing it has in mind is machine learning to boost IoT system efficiency.
NVMesh can be run in grouped server nodes to provide external storage to accessing hosts, what Excelero calls a converged infrastructure, as well as in hyperconverged designs.
NVMesh is NVMe over fabrics ready, and much of the need for RDDA may go away when NVMe over fabrics becomes a reality.
Dell has signed a reseller agreement with Excelero; yes, the Dell that recently canned its DSSD NVMe-based array. Imagine ScaleIO running in an Excelero-based hyperconverged system.
Clearly management facilities at the 100,000-node level will need to be developed, such as to understand and deal with hotspots and how to monitor and manage overall and per-component performance. We understand that, at this early stage, Excelero does not have a data center bridging capability. ®