Data Locality

StarWind NVMe over Fabrics
(NVMe-oF) Target. Build the fastest
storage ever

We all know that IT environments built for high-performance computing (HPC), stock trading, databases, or whatever else IOPS-hungry can never have too much performance. That’s actually why you typically need an all-flash storage array to run any of those applications. Think big! Use NVMe SSDs – the true king of the hill when it comes to performance.

Problem

Of course, these days, it’s not a big deal to build an all-flash environment with one or two NVMe drives on board. An infrastructure like that still costs you a dime, but, at the end of the day, you always get the performance that you are looking for. Well, what if we tell you that your applications are likely to access only half of all-flash storage performance?

The problem is in protocols that are used to talk to SSDs. All your applications use legacy protocols like iSCSI, iSER, SMB3, and NFS to access the underlying storage. They were designed to connect slow spindle drives, not flash! The problem is, these protocols have a single short command queue. That’s, actually, why they entail significant I/O overhead once being used with NVMe SSDs

Serial Attached SCSI (SAS) – good for spindles, but far too slow for flash

That being said, it seems obvious why your performance-hungry applications do not get the true NVMe SSD capabilities while using legacy protocols. Of course, adding flash to your setup usually grants you like 20% more IOPS than spindle drives can provide. But, let’s face it, it’s nothing compared to the true SSD performance! NVMe needs a whole other tech…

Solution

NVMe over Fabrics (NVMe-oF) is the protocol allowing to unleash the true NVMe potential. The single short command queue is replaced with 64 thousand queues, 64 thousand commands each. Such design enables NVMe devices connected over NVMe-oF to achieve the consistent microsecond-scale latency and allows you to squeeze all the performance that flash can potentially provide. In our lab, we managed to get over 2M IOPS out of 4 Intel Optane SSD 900P drives in a bare metal environment. And, we are heading beyond!

The traditional way to access underlying storage VS. NVMe-oF

But, only Linux hypervisors feature native NVMe-oF support. Other industry-leading hypervisors like Microsoft Hyper-V and VMware ESXi leave users to old protocols which are proven to be inefficient for flash. No fun!

To bring NVMe-oF to all hypervisors, we added a new feature to StarWind Virtual SAN – StarWind NVMe-oF Target. It is a Storage Performance Development Kit (SPDK) component designed to run the whole data path from the network to a flash device completely bypassing the kernel. Wait, no SPDK in Windows? No problems, we bring the Target in a VM where OS can do SPDK!

StarWind NVMe-oF Target in the nutshell

Conclusion

When it comes to storage performance, NVMe is the true king of the hill. However, it is still challenging to squeeze all NVMe SSD IOPS as traditional SCSI-derived protocols do not work that good for flash. The solution is StarWind NVMe-oF Target – StarWind Virtual SAN feature bringing blazing fast NVMe SSDs to businesses.




Request a Callback Gartner`s Niche Player