Are you ready for network democratization?
That’s the goal of a new networking approach: NVMe over TCP with ADQ. Simply stated, this networking combo promises the higher speeds of networks purpose-built for storage, but without the complexity and higher costs.
NVMe — short for Non-Volatile Memory Express — is a transport protocol for storage over PCIe. It offers low latency (aka high speeds) with high reliability. But to transport NVMe over a network, you also need to protect the data from interruptions and dropped packets.
Engineers have tried to overcome these weaknesses with purpose-built networks for storage. These include Omni-Path, InfiniBand and Fiber Channel.
But implementing these alternative setups requires skills possessed by very few IT shops. Also, these purpose-built networks can be costly.
Hidden in plain sight
A better solution made use of a network protocol many organizations already had in place: Ethernet.
Ethernet has a lot going for it. For one, it’s a general-purpose network. For another, it’s been around for a long time, so the needed skills are easier to find. Ethernet is also good at recovering from errors, interruptions and dropped packets.
But there’s a catch: Compared with purpose-built networks, Ethernet networks have a lot more latency. In other words, they’re slower.
To overcome these limitations, engineers next came up with 3 Ethernet-based alternatives: ROCE (pronounced “rocky,” it’s short for RDMA over Converged Ethernet); iWARP; and NVMe over TCP.
How well did these work? Well, the following chart, courtesy of Intel, shows how these approaches compare on 2 important measures: IOPS (input/output operations per second) and latency, the time needed for a packet to travel from source to destination.
As you can see, NVMe/TCP (green bar) was the worst performer on both:
This was the problem Intel set out to fix. Ethernet and TCP technology were already familiar to the masses. Now how could these network standards be made to perform better?
The solution? ADQ, short for Application Device Queues.
The Intel Ethernet 800 Series has Application Device Queues (ADQ) that act as dedicated express lanes for high-priority applications. ADQ filters application traffic to a dedicated set of queues for predictable high performance.
Essentially, ADQ accelerates Ethernet network performance. This provides more predictable application response times, lower latency and higher throughput.
So how does this new approach perform? The following chart, also courtesy of Intel, tells the story. NVMe over TCP with ADQ (the gold bar to the right) offers high IOPS with just half the latency of NVMe over TCP without ADQ:
While the performance of NVMe over TCP with ADQ acceleration doesn’t yet beat that of the RDMA solutions, it does get on par. So your customers can take advantage of the ease-of-implementation and scalability of NVMe over TCP — and without feeling a large performance impact.
Intel isn’t keeping this solution to itself. The company has made ADQ open source through updates to the Linux kernel. Also, ADQ support is built into Intel’s Ethernet 800 Series network adapters.
Intel Ethernet 800 series products started shipping to select customers in late last year. Intel says broader availability is set for this year’s third quarter.
The bottom line
NVMe over TCP with ADQ gives you and your customers the ease of NVMe, the familiarity of Ethernet, lots of flexibility and ease of use. And it uses standard NICs, standard Ethernet switches, and standard network configuration.
NVMe over TCP with ADQ also eliminates the need to reconfigure a purpose-built network for storage. That can save your customers both time and money.
Sound good? That’s the democratization of networking.