The DevOps world is ripe for a storage paradigm change, both in how it operates within the enterprise and how it implements applications.
NVMe, especially NVMe-over-fabrics (NVMeoF), is both a tool and a host target that can provide such a change—especially in the new world of microservices and containerization. NVMe can drastically improve application performance over classic implementations and, given the right tools, transform the process of writing those implementations to be faster, more robust and more flexible.
A quick recap shows why this is so. The DevOps build toolchain, especially for large codebases, has always been heavily I/O-dependent and storage access has classically been the biggest bottleneck for building an image. The ability to cluster small, inexpensive Linux x86 servers connected to SANs twenty years ago was a first step in separating storage hardware from compute, but I/O latency was pretty awful. Build/link operations could actually take longer than the compilation of object files.
Virtualization helped make hardware utilization better and cheaper but didn’t really improve throughput. Cloud infrastructures such as OpenStack and CloudStack made those simpler infrastructures and more elastic to manage but, again, the IOPs bottleneck remained.
Containerization for DevOps made it even simpler from the applications level—just deploy the app and ramp up container instances as needed—no more dedicated infrastructure needed. And it’s cheaper for incremental functions.
However, not much was done to clear the I/O bottlenecks for largescale storage and increase overall performance.
When the new nonvolatile memory technology of NVMe came to market in 2014 for the single server world, it delivered two game-changing capabilities.
- An increase of 5x disk access I/O speed. This was, arguably, the biggest single performance improvement since multicore processors with onboard cache (think Opteron) emerged. With Linux adding support for standard, open source NVMe drivers in 2021, the options increased beyond Windows and Mac users.However, SANs and their protocol stack were still an issue. SATA- and SCSI-based SANS performance was gated to the 1.2GB limitations of those technologies and big iron to head-end the SAN still dictated high-cost infrastructure.
- Unlike SATA or iSCSI, NVMe has the native capability for TCP/IP or fiber channel addressing. This allows NVMeoF to be an autonomous point for data storage in a normal network. A user can now keep control of their data but use a network-attached compute as a utility with roughly 6Gb speed for I/O—greatly lessening the big penalty for disk I/O.
The dramatic uptick in container deployments and the debut of container attached storage (CAS)—software that includes microservices-based storage controllers that are orchestrated by Kubernetes—represented another industry milestone. These storage controllers can run anywhere that Kubernetes can run, which means any cloud or even bare metal servers or on top of a traditional shared storage system. Critically, the data itself is also accessed via containers as opposed to being stored in an off-platform, shared scale-out storage system.
Since container-attached, cloud-native storage enables the use of small, disaggregated architecture without the need for large-scale SANs or storage controllers dovetails nicely with DevOps work.
Container-based applications can benefit enormously from NVMe-based storage. As experts have pointed out, since DevOps applications are all about scalability, the parallelism embedded in the NVMe spec fits nicely to support them.
While SANs are good for large-scale contiguous operations, they’re expensive chunks of capacity and not nimble to deploy. They’re also bound to the slower I/O storage protocols, which can negate performance gains possible by adding NVMe to that array (as opposed to discretely).
Container-based storage with NVMe gives DevOps teams the opportunity to add storage at low cost with deterministic management of that stateful storage—in theory, of course. In practice, the temptation to add like infrastructure to current infrastructure is powerful, even with all its limitations.
To make that theory a reality, a control plane with the fast data plane is necessary—and the market is beginning to address this need.
As one example, OpenEBS recently introduced an upgraded engine, MayaStor (now part of DataCore Software), which enables use of NVMeoF at impressive speed: Benchmarks with Intel SPDK architecture show a 5x read/write speed improvement over SATA.
OpenEBS debuted a few years back as a containerized storage solution for Kubernetes-based applications. With an optimized storage engine, Kubernetes has the opportunity to perform like other, more specialized hosting and development solutions.
Impact for DevOps
What does this mean for DevOps teams? Firstly, using NVMe in containerized DevOps applications speeds up the time required for large builds, thereby speeding iterations of writing software. It also removes the need for complex and expensive SANs to host build environments. This double whammy of benefits means that development is quicker and cheaper and, in turn, applications can deploy faster while using more flexible tools.
Second, NVMe greatly mitigates I/O concerns for how application code is designed and how environments are implemented. No more being beholden to the storage trunk—NVMeoF drops the performance penalty for I/O ops and distributes traffic across a normal TCP/IP network. This is now a triple whammy—faster, cheaper and more flexible containers built atop a fast, flexible virtualized infrastructure.
Backed by innovations in container-attached storage, DevOps teams can achieve the agility and flexibility that are paramount to its mission—and act as a proving ground for the rest of the IT team to consider alternative paths to greater efficiency and performance.
Join us for KubeCon + CloudNativeCon Europe 2022 in Valencia, Spain (and virtual) from May 16-20—the first in-person European event in three years!