Part of:Best practices to manage NVMe performance and power
4 NVMe storage problems you can prevent before they happen
Implementing NVMe storage is complicated, from knowing what to ask when buying NVMe SSDs and arrays to what type of file system to use. Here's how to deal with four NVMe issues.
NVMe is rapidly taking hold as enterprises seek to capitalize on its performance and low-latency advantages. But NVMe storage problems are emerging as well.
NVMe SSDs make up more than half of all enterprise SSD shipments, according to G2M Research. In addition, all the major all-flash array vendors now offer NVMe- and NVMe-oF-based AFAs, G2M said, and the total NVMe ecosystem is expected to reach $65 billion by 2021, according to G2M's estimate.
The storage industry developed NVMe from the start for SSDs, with an emphasis on improving throughput and IOPS and reducing latency. Because of this, the protocol helps overcome storage infrastructure bottlenecks that resulted from using SSDs with legacy interfaces, such as SATA and SAS, which were developed for HDDs. But, along with the advantages, come complications as the technology reveals weak links in other parts of the storage infrastructure. NVMe's full effect on the enterprise is just now starting to be felt -- both the good and problematic.
Be aware of potential issues as your organization moves along the path toward implementing NVMe. What follows are some NVMe storage problems to watch for, how you can keep them from happening and how to tackle them if they do.
NVMe's full effect on the enterprise is just now starting to be felt -- both the good and problematic.
NVMe SSDs differ in endurance, I/O consistency and quality of service. And while prices are falling, a low price can also mean lower endurance or performance. Check the fine print, and ask a lot of questions. With arrays, pay attention to what sort of management software is included, what level of NVMe support you're getting and how that will affect performance.
In addition, check that any product you're considering is fully NVMe-compliant. For example, some mostly proprietary products are being passed off as compliant but may not deliver expected performance, endurance and cost savings. Make sure any product you're considering meets NVMe specs and has passed NVMe compliance tests.
When it comes to NVMe AFAs, performance improvement depends on the architecture used. Conventional AFAs using the NVMe protocol generally improve performance by about 20%, but emerging designs should deliver a much higher performance boost.
This NVMe performance gap exists because not all AFA controllers unleash the full performance benefits of NVMe SSDs. Enterprises can get optimum performance by using AFAs that provide end-to-end NVMe connectivity. These arrays use NVMe-oF to connect to hosts via Fibre Channel (FC) and Ethernet networks at the front end, and they provide back-end connectivity to NVMe SSDs. This approach enables hosts to use the native NVMe protocol to talk directly to the NVMe SSDs and send native NVMe commands over the Ethernet and FC networks that pass through the array's controllers.
Composable infrastructure is also starting to provide other options for enhancing NVMe performance.
Don't let file system structure drag down NVMe storage
One problem with NVMe is it can reveal weaknesses in other parts of the storage infrastructure. Any weak link will end up increasing latency and counterbalancing NVMe's latency efficiency.
The file system is one such weakness that can cause a bottleneck. Most NVMe storage systems are designed for block storage, circumventing the performance issues of the file system. But most modern applications -- such as AI, machine learning and data analytics -- need a file system, so often one is added to the block storage. With traditional file systems, all I/O goes through a primary node that high-velocity workloads can easily overwhelm.
There are two types of file systems to get around this issue. Look for ones that write directly to NVMe drives instead of through the OS I/O stack. Alternatively, look for ones that communicate across NVMe-oF so they manage metadata more efficiently.
Get more details about file system design and how to keep it from becoming an NVMe storage problem.
Dealing with the NVMe storage management challenge
Once you have all those NVMe SSDs and arrays in place, you're going to need a way to manage them. Fortunately, the recently released NVMe Management Interface (NVMe-MI) specification provides an infrastructure for managing NVMe devices, enabling direct out-of-band and in-band administration of NVM subsystems. The spec defines a command set and architecture for controlling NVMe storage. It enables remote management applications to discover, monitor and update NVMe devices.
NVMe-MI doesn't require any specific NVM storage, system processors or OSes; it also doesn't recommend any specific usage models. It aims to provide a common management interface that abstracts the implementation details. NVMe-MI infrastructure is comprised of four layers -- application, protocol, message transport and physical -- that work together to handle management-related communications.
