Disaggregated Storage Part II with Zivan Ori from E8 Storage

In this podcast, originally recorded on 30 October 2017, Chris and Gavin follow up on the subject of disaggregated storage with Zivan Ori, CEO at E8 Storage. E8 is a start-up that markets and sells a storage array that uses a disaggregated architecture.

The conversation covers what disaggregation really means and how E8 focuses on splitting the data and control paths of traditional I/O commands. With NVMe and RDMA, the E8 architecture allows direct access to metadata on dual controllers within the E8 hardware. This piece was actually a revelation during the conversation as it explains how shared direct access to NVMe drives can be achieved without each connected host getting in the way of another.

The conversation wraps up with a discussion on typical customers and what applications need this level of high performance and low latency. The original podcast on disaggregation can be found here – Garbage Collection #005 – Disaggregated Storage.

Note that this podcast is a little longer than our normal shows, however it is packed with some great insights into the use of NVMe versus SAS/SATA drives and much more.

Elapsed time: 00:41:36

Timeline

00:00:00 – Introductions
00:01:00 – Disaggregated Storage recap
00:02:00 – Separating the data and control paths, why this matters
00:03:30 – How much better are NVMe drives than SAS/SATA? (Including Zivan’s laptop)
00:05:30 – What about the price differential – are NVMe drives more expensive per IO?
00:06:53 – Apologies for the background sneeze….
00:07:23 – Betting on commodity trends – not taking the DSSD route (or Violin)
00:08:20 – Getting into more technical detail
00:10:00 – How can you break the dual controller model with just software?
00:10:52 – Using RDMA to extend the control and data planes outside the shared array
00:11:36 – How important is networking in this architecture?
00:12:35 – Fibre Channel doesn’t cut it any more?
00:14:06 – Why not build a server solution, rather than a storage solution?
00:15:48 – Shared storage still provides an advantage over hyper-converged
00:17:00 – Getting into E8 architecture in more detail
00:18:52 – Client requirements, drivers, agents and emulated block devices
00:20:00 – The secret sauce – splitting the data and control paths for generic I/O commands
00:25:20 – and here’s the key – being able to talk to the metadata servers with very low latency
00:27:00 – New architectures – driven by RDMA and NVMe
00:28:48 – How is data protection within the box implemented?
00:32:30 – What does a typical customer look like?
00:38:00 – How does faster storage lead to better business advantage?
00:39:17 – What’s next?
00:40:50 – Wrap up time