Several NVMe flash vendors are claiming that NVMe Flash has reached price parity with Serial Attached SCSI (SAS) based flash. The claim is true or pretty close to true in most situations. Vendors are selling their NVMe flash drives for the same price as SAS based drives. The problem is that to achieve the full performance potential of the NVMe investment requires an ecosystem of components and the vendor or IT will need to upgrade each of them. The cost to upgrade each of the components of the ecosystem can significantly add to the total cost of an NVMe solution, and parity is not yet within reach. As a result, IT needs to deploy NVMe only on workloads that can fully exploit its low latency.
Understanding NVMe’s Impact
NVMe impacts the rest of the ecosystem because its low latency exposes bottlenecks. In the past, other components within the storage ecosystem could hide behind the high latency of hard disk drives. SAS Flash started to expose other bottlenecks but NVMe flash makes it readily apparent. To compensate, vendors are adding multicore processors, more RAM, field gate programmable arrays (FPGAs) and custom application-specific integrated circuits (ASICs). All of these additions add to the solution cost.
The second impact is NVMe’s ability to network via NVMe over Fabrics (NVMe-oF). A flash array with internal NVMe flash drives and combined with external connectivity via NVMe-oF, promises to deliver latencies that rival internal server flash. The problem is that most NVMe flash arrays do not, support external NVMe-OF connectivity. IT planners that want the lowest level of latency, but still want the efficiency of a shared storage system, need to select an NVMe Flash Array with external NVMe-oF connectivity. They also need to make sure they are running the latest network switches and network adapters. Most switches made within the last year should only require a software update. The network infrastructure is slow to evolve and it is quite possible that the majority of switches and network adaptors in the data center are not ready to transport NVMe so they will require an upgrade.
Most data center upgrades happen when the cost of the new technology is the same or less than the current technology. For NVMe this means a combination of better storage software optimization so that less proprietary FGPAs/ASICs and more powerful CPUs are not needed. It also means it will take time for networking infrastructure to gradually be upgraded so that is supports NVMe-oF natively.
What Should Data Center’s Do About NVMe Right Now?
The cost of the surrounding ecosystem does not mean organizations should wait until NVMe is truly cost neutral. Instead they need to look for opportunities where investing extra or accelerating upgrade plans will pay off. Many organizations will have a handful of workloads that will benefit from an immediate move to NVMe. The challenge is in understanding how and when they should make that investment. In most cases the potential applications or workloads will, if performance can be accelerated, provide the organization with the potential to increase revenue, improve customer satisfaction or lower costs through deeper infrastructure consolidation.
How though, does the organization decide which applications qualify? They can look at benchmarks and vendor claims about the hardware but those data points don’t tell the organization how their application will perform on the new hardware. IT can create a lab that attempts to simulate production but it is difficult and expensive to gather all the components needed for a proper simulation.
An alternative is to use workload modeling to create an exact model of the application and replay that IO on any new system that the organization is considering. Workload modeling works not only on testing the potential improvement of NVMe systems but also on other scenarios like moving a workload from Fibre Channel to iSCSI or from NFS to Object Storage.
NVMe in servers has reached price parity and is worth it for those workloads that need internal, direct attached storage. The more common shared storage infrastructure still has to be further optimized, primarily through storage software optimizations, before NVMe makes sense universally. Today, many organizations can meet their performance needs easily with a SAS based all-flash array or even a hybrid (flash and hard disk) array. IT planners need to be careful where they place their NVMe bets. The investment has to be on applications that can take advantage of the technology’s low latency and those applications have to show direct positive impact on revenue or customer satisfaction as a result of the performance gain. Workload modeling is the most logical way to make sure the organization’s NVMe investment bet pays off.
In our 15 Minute on demand webinar, “NVMe Readiness Assessment” Storage Switzerland and Virtual Instruments take an objective look at NVMe and help you map out a strategy for the transition. Watch this rapid fire discussion to learn if NVMe is right for you and if you are right for NVMe.