Non-volatile memory express (NVMe) flash drives promise new levels of ultra-low latency and fast performance. The media is beginning to find its way into the enterprise storage environments, on the back of the need to serve the growing number of performance-intensive workloads.
NVMe storage is typically deployed on a direct-attached basis as opposed to through a storage area network (SAN) to maximize performance. However, this creates a problem whereby the server’s central processing unit (CPU) resources quickly become overwhelmed by the levels of input/output operations per second (IOPS) that NVMe storage media enables (most are capable of 750,000-800,000 IOPS). The customer is forced to purchase more storage drives and controllers – an expensive problem that quickly compounds itself, considering the density of NVMe flash drives (we will see drives capable of 16-32 terabytes (TB) of capacity hit the market soon). Josh Goldenhar, Excelero Vice President of Products, recently joined George Crump, Storage Switzerland Lead Analyst, to discuss how his company’s modernized SAN architecture enables enterprises to harness the full performance and capacity density advantages of NVMe.
Excelero’s NVMesh architecture pools local NVMe storage resources across any storage networking, including Ethernet or InfiniBand. This frees available storage capacity from being bound to individual servers, and as a result improves capacity utilization. Excelero’s Remote Direct Drive Access (RDDA) driver then intelligently creates logical virtual volumes of storage and exposes those volumes to applications as local block storage devices. Through this architecture, the application can access the read performance of all pooled drives and the write performance of all drives within the logical volume – facilitating processing speeds and latency levels that are on par with local storage, according to Excelero. It also adds redundancy to protect against a host or drive failure. Because there is no centralized controller, the architecture furthermore can scale easily, potentially to multiple petabytes (PB), while avoiding performance bottlenecks.
Storage Switzerland recognizes that NVMe over Fabrics (NVMe-oF) implementations are emerging to facilitate shared NVMe storage with limited added overhead. The comparative value add of NVMesh lies in its ability to also handle logical volume pooling, mirroring and data striping – offloading these tasks from the host CPU. Offloading these data services tasks from the CPU furthermore enables them to be scaled more easily across the environment, and helps to improve application performance predictability. According to Excelero, NVMesh adds only approximately five microseconds of read latency, still enabling a read to be conducted across a logical storage volume in approximately 85 microseconds.