Jul 1, 2016

E8 Storage promotes NVMe over fabrics

E8 Storage (www.e8storage.com), innovator in low latency storage, has driven and delivered an interesting SNIA tutorial "High Availability for Centralized NVMe" during the last DSI conference mid June. Zivan Ori, CEO and co-founder of E8 Storage, did a tough exercice to cover a topic and not to do any promotion for its own product that play in this segment, aligned with SNIA rules to present a neutral tutorial.
The idea of this talk was to introduce the need for HA when you consider a centralized NVMe storage array, like in the past SAN storage array, but here with the constraint of the latency, that changes the solution.

Zivan Ori, CEO and co-founder, E8 Storage
NVMe drives, introduced in 2014, has very limited deployment models usually embedded in servers or laptops in a DAS classic way. In that case, NVMe drives are super fast with very low latency. Even with some problems and challenges like redundancy and resiliency to name one key point, NVMe is still very interesting especially when you consider sharing drives between servers. But now, when you do that, the remote latency diverges from local one I mean when the drive was locally attached. Second point is about pooling NMVe drives in a shared array with the bandwidth/throughput explosion challenge. The obvious idea would be to connect a NVMe array via a SAN but immediately we found 2 problems with the I/O stack and the storage network. For these reasons the industry has worked, specified and published in June an official specification of NVMe over fabric aka NVMf. The key component here is RDMA available on IB, Ethernet named ROCE (RDMA over Converged Ethernet) or TCP/IP. But this NVMf approach solves only half of the problem only the latency challenge with remote NVMe drives but not the resiliency and other features represented by the I/O stack. To solve that, 2 choices exist:
  1. implement data services on the host with a sort of very thin storage agent. It allows scale-out offering with a SDS model with features implemented outside the array, the logic is running on the host. But some questions exist such coordination between agents et still the resiliency of the platform. But among the 2 modes, this one is the most scalable of course.
  2. run these in the controller itself that make possible only scale-up products pretty similar to "classic" controller-based storage appliance and as mentioned you scale "behind" the controller.

The page 25 of the presentation below shows a good table that compares the 2 modes. Data Centers are evolving and NVMe and NVMf are a reality with different offerings on the market, each promoting its own value and differentiator. If you wish to read some vendors collateral, go to Apeiron Data Systems, DSSD, E8 Storage, Excelero, Kaminario, Mangstor or Mellanox.

1 comment :

Baruch Even said...

One point was omitted in this presentation, it is now the responsibility of the customer to provision the compute units with enough compute power to handle the added workload of doing the RAID and data services on top of the actual application need. It is not a big drawback and depending on the usage might not even be a drawback at all but it should be mentioned and remembered. All in all this sounds like a very interesting product.