QSAN Blog

NVMe Storage Matters for Database, AI, and HPC

blog-NVMe Storage Matters for Database, AI, and HPC

Table of Contents

The Evolution of NVMe Storage

As flash-based SSDs (Solid State Drives) gradually become mainstream and getting to replace traditional hard drives, the SAS (Serial Attached SCSI) and SATA (Serial Advanced Technology Attachment) protocols are no longer sufficient to provide enterprises with the speed required to fully utilize their SSD storage. In this way, a newer storage protocol NVMe (Non-Volatile Memory Express) developed specifically for non-volatile memory will become an important part of many data centers in the future. Analysts such as IDC believe that NVMe storage will replace traditional storage protocols, especially for major workloads that are sensitive to latency.

Speed Processing with Faster Storage

SSD has made great progress on disk drives, changing the data access latency from the 5ms average seek time delay of disk drives to 200μs (0.2ms) of SSD. After adding enterprise features of dual ports and reasonable endurance, SSDs enter the array market in the form of AFAs (All-Flash Arrays), also as caches and tiers in hybrid disk arrays. Although hybrid drives are slower than AFA, they have more data and are faster than full disk arrays.

NVMe greatly improves the speed of SSD again. This allows them to directly access the PCIe bus, reducing the latency to 30μs (0.03ms). This helps the server to access directly connected NVMe SSDs. NVMe storage is able to do this because its command set requires less than half the number of CPU (Central Processing Unit) instructions to process I/O requests compared with the command sets of SCSI (Small Computer System Interface ) and ATA (Advanced Technology Attachment). NVMe supports 64K commands in a message queue and up to 64K queues. Compare to the traditional protocol, SAS devices only support up to 256 commands per queue, while SATA supports up to 32 commands.

Latency Matters Your Business

NVMe storage was originally a host controller interface designed to use the PCIe bus to quickly transfer data between the host and the target SSD. The goal is to reduce the transmission components in order to reduce latency, improve performance, provide parallel I/O, and even reduce power consumption. These advantages are very suitable for the following use cases using NVMe flash storage.

  • Database: NVMe storage has quickly become the first choice for companies using intensive relational databases. The enhanced performance of NVMe flash storage systems reduces the number of physical servers and database licenses required.
  • AI (Artificial Intelligence): Today’s AI and machine learning applications are all about speed, processing data much faster. They need to be superior to the past. They also rely on larger data sets, especially for training intelligent system algorithms. NVMe provides the bandwidth and low latency required for these demanding workloads, making it a mainstream choice for AI storage.
  • HPC (High-Performance Computing): NVMe SSDs can be used for HPC or FinTech (Finacial Technology) applications, such as high-frequency trading, where low-tech latency is essential. It can speed up calculations by reducing the time the processor waits to read data from storage.
  • Virtualization: Enabling more VMs (Virtual Machines) on physical servers is another way NVMe storage can benefit enterprises. Virtual systems with a high ratio of virtual machines to physical servers have a lot of I/O load. NVMe can support the I/O capabilities needed in order to meet these needs.

Higher IOPS and low latency are related to your business repeatedly, of course, this will also affect your IT budget.

Next, Improved Storage Network Fabric Communications

NVMe is more than just a high-performance host controller interface for direct connection to SSD. It has developed into storage architecture. With the emergence of structure-based NVMe and NVMe-oF (NVMe over Fabrics), IT administrators can use common storage structures to extend the advantages of NVMe storage on their IT infrastructure.

NVMe-oF defines a common architecture that supports a series of storage networks for the NVMe block storage protocol on the storage network. This includes enabling front-end interfaces to storage systems and expanding to a large number of NVMe devices. NVMe-oF supports Fibre Channel, RDMA (Remote Direct Memory Access), and TCP/IP protocols. The benefits of NVMe-oF include faster connections between storage and applications on the server and more efficient CPU usage. This can further integrate data center and network usage.

Conclusion

Now, data has become a valuable asset in this era of digital transformation. NVMe technology can eliminate the performance gap between memory and disk drives, thereby opening up a new model for building its applications and services.

The storage advantage of NVMe is that the technology has fundamentally changed the way how IT administrators view storage and memory. NVMe no longer regards storage as a place to put data, but turns storage into an extension of memory, thereby eliminating extra steps and inefficiencies.

This feature is very important for applications that require ultra-high performance. Achieve the same application capabilities, but use less equipment. It is also very helpful to the business to reduce the TCO (Total Cost of Ownership).

Official Blog

Latest Trends and Perspectives in Data Storage Management