The primary goal of the NVMe standard is to reduce latency. It accomplishes this by leveraging the PCIe interface and increasing both command counts and queue depth. NVMe Flash Systems however, are more than just the flash media. Vendors create these storage systems out of computing resources, networking resources, and storage software to control how data flows to and from the NVMe flash storage. The NVMe flash is a component within an ecosystem that creates a storage array. Each of the components can impact latency, potentially negating the latency reductions that NVMe offers.
The Storage Software Challenge
A primary point of concern is the storage software itself. The first challenge is the efficiency of the storage software. The software needs to be optimized for today’s modern multi-threaded processors. The features it provides must also be efficient so that the capabilities offered by NVMe flash to reduce latency, which enterprise IT has come to count on, can be fully realized.
Vendors have taken several approaches to work around the latency introduced by the storage software. One of the more common is to offer featureless storage software that only provides the most basic storage functions like volume creation and management. These vendors’ stance is that modern applications such as Hadoop and NoSQL databases have the most common storage features built in, so IT doesn’t need those features in the storage software. The problem with this approach is that most data centers have not cut over 100% to modern applications and most never will be 100% modern applications.
The second challenge is forcing the application to perform storage management functions and deliver storage features. This means the featureless array diverts some CPU power, intended for the processing of the application, to storage management functions. The organization ends up over-buying computing power for the application, and it is much more difficult to deliver consistent performance.
The third challenge is that while it is reasonable to assume that feature-rich storage software adds somewhat to latency, IT planners can’t ignore the reality that most organizations can sacrifice some latency in return for a richer feature set and never notice a performance impact.
Striking a balance – Features and Low Latency
For most organizations, the better choice is to sacrifice some latency in return for full-featured performance. The storage vendor needs to optimize its software for today’s multi-threaded processor designs and make sure their features such as deduplication, compression, snapshots, and replication impact performance as little as possible.
A full-featured NVMe flash array might introduce some latency, but the organization can quickly realize a gain by its introduction. First, the computing power dedicated for the applications can remain dedicated. Second, most of the features in a new NVMe all-flash array are there to either protect data from media failure or to reduce operational costs. If the organization is using a featureless NVMe flash array and counting on the application to provide the missing features, they often see an increase in application processor consumption and in flash capacity consumption since most applications use three-way replication instead of a RAID-based data protection method.
Finally, a featureless NVMe flash array might promote additional storage sprawl since it more than likely is a block-only device. This type of NVMe flash array also needs a full-featured file system to support artificial intelligence, machine learning, and high-velocity analytics. A full-featured system may also be multi-protocol, which means it can provide an NFS file system for AI and ML workloads while at the same time being the storage solution for more traditional workloads such as Oracle and VMware through either NFS, iSCSI or Fiber Channel (FC) protocols.
In our next blog, we discuss the role of Serial Attached SCSI (SAS) in the era of NVMe. Check in next week to find out if there is a place for SAS flash in the modern data center. While you are waiting for it, sign up to join experts from Storage Switzerland and Western Digital for our live webinar, “Myth Busting – The Four NVMe Myths.” In it, we will discuss the top four NVMe myths and explain why they should be “busted.”
Watch the webinar and get answers to these questions:
- Can “regular” data centers benefit from NVMe or just companies specializing in AI?
- Does NVMe have to be end-to-end for data centers to benefit?
- Do features such as deduplication and compression impact performance?
- Does NVMe Flash mean that SAS Flash is no longer relevant?
Join out on demand webinar to get the answers to these and other NVMe questions. All registrants receive a free copy of Storage Switzerland’s latest eBook “Busting the NVMe Flash Myths”.