NVMe over TCP and Why it Matters

It’s been a few months now since I joined Fungible, so I thought it was about time to offer some initial thoughts about my experience so far. Let me start off by answering a common question I get, “Are you still glad you joined?” That’s an easy one – Absolutely! It’s a great company with incredible people and a culture of true innovation. I’m more convinced day by day that we are poised for massive growth. After a couple of months, some of the shine can wear off in a new job as you start to see deeper into the day-to-day operations. Fungible is no exception – there are certainly areas we are working on. One of those areas is being really clear about what we provide to the market. This will directly lead to growth opportunities, because what we have to offer is really tremendous.

So in service to that goal, let me share one of the key areas where we are working to clarify our message, and why it matters. Fungible offers NVMe over TCP (NVMe/TCP) scale-out storage. We’ve referred to that more generally as NMVe over Fabrics in the past, but the way to access our storage cluster is using NVMe/TCP. That’s what we were designed for all along, and that’s what we’ve supported since we launched the Fungible Storage Cluster (FSC) a year ago.

First off, NVMe/TCP is based on standard IP over Ethernet networks. That means the network you have already deployed in your data center is ready to take advantage of our storage. No need to build out a separate storage network or to buy special switches or network adapters to take advantage of the performance and scalability we offer. Just plug it in and go!

This leads to the second reason NVMe/TCP matters – we can scale along with your network. Because FSC is plug and play on TCP/IP ethernet, you get to decide where within your data center you want to place our storage nodes – in one rack, one per rack, one per row. You decide based on the needs you have. We support top of rack network speeds of 10Gb, 25Gb, and 100Gb. So you can scale your storage performance as you scale the speed of your network!

Thirdly, NVMe is really fast over the network. I mean REALLY fast. Why is it so much faster? NVMe/TCP removes the translation layer from Drive Protocol to Network Storage Protocol. It lets you put the Drive Protocol directly on the network. Removing this translation layer means every single communication back and forth saves 2 translations! The Fungible Storage Cluster further optimizes the data path by removing the general purpose OS and CPU from the mix entirely. Bare NVMe performance directly across the network from the storage to the host. Many vendors support specific implementations of NVMe over Fabrics to try and optimize the speed of NVMe (like RoCEv2), and claim that NVMe/TCP is slow. But the reality is, NVME/TCP is really fast when you build the processing and storage solution around that protocol to begin with, like we did. In fact, we are the only engineered (optimized in silicon, in our Fungible DPU™) high performance storage product on the planet that I am aware of built from the ground up to run on NVMe/TCP.

By the way, don’t just take my word for it. Nikhef recently validated this performance (https://www.nikhef.nl/en/news/22705/) and stay tuned for more exciting news to come!

There’s so much more to unpack about Fungible and the really amazing capabilities the team here has built, but I can’t cover it all in one blog. Stay tuned for my next installment. And if you are interested in high performance and massively scalable storage that will drop right into your existing datacenter network to support your mission-critical data storage and analytics needs, give us a shout! We’d love to talk with you about what we can do for you.

For more information, or to set up a conversation with us, contact [email protected].

2 thoughts on “NVMe over TCP and Why it Matters

Comments are closed.