Shulou(Shulou.com)06/02 Report--

NVMe enables the storage infrastructure to take full advantage of flash memory by improving the physical interface, increasing the number of commands, and queue depth. But NVMe also presents a challenge: the latency of NVMe is very low, which exposes weaknesses of other components in the storage infrastructure. Any weakness in the infrastructure increases latency and reduces the value of NVMe.

The file system is a big problem in the storage infrastructure. It is time for vendors to rethink the file system architecture, in particular, they must modify the way the file system interacts with NVMe storage to avoid becoming a bottleneck.

Photo by panumas nikhomkhai from Pexels

Why the file system?

File systems that serve AI and high-speed workloads are usually scalable. A scale-out file system consists of multiple storage servers or nodes, and the file system aggregates internal storage in these nodes and represents it as a single storage pool that users and applications can access. Traditional file systems can also be scaled out, but they are serial, which means that all iUnix O passes through a master node, which can be easily overwhelmed by AI and high-speed workloads, creating bottlenecks. These workloads rely on a parallel file system structure that allows any node in the cluster to provide IWeiO services to users or applications, which makes network efficiency even more important.

Most NVMe storage systems are designed for block storage. As a result, they avoid the performance overhead of the file system architecture. In most cases, however, the file system is added to the block storage system so that it can be used by these AI and high-speed workloads. Most modern applications-- especially AI, machine learning, and big data analysis processors-- rely on file systems.

A well-designed, block-based NVMe storage system with the addition of a file system may still be faster than a block-based SAS storage system, but the performance degradation between RAW block storage and file system-controlled storage is significant. Therefore, the organization needs a file system optimized for NVMe.

What should you focus on for benchmarking

Vendors often use several file system benchmarks to demonstrate their functionality. Most of these tests use NVMe block storage with parallel file systems, and vendors can easily use various configurations to bring a parameter to the top of the chart, which can be misleading.

For example, in the current standard performance evaluation company SFS 2014 benchmark, top vendors differ significantly in the number of drives, drive types, and the number of storage nodes in the test environment. In most cases, hardware vendors try to reduce file system architecture overhead by using hardware that exceeds requirements and push prices beyond a level that is reasonable for most organizations.

What really matters is the implementation of the hardware and file system on the workload type and budget of the organization, and most companies do not have unlimited funding to create a perfect combination of NVMe- file systems. IT professionals should look for the simplest configuration that meets the requirements.

What should be concerned about the file system

There are three main limitations to file system performance:

Communication efficiency between file system and storage node

How the file system efficiently manages the network that connects to each storage node and how to communicate with the client efficiently

How the file system effectively manages metadata access.

In most modern application environments, metadata accounts for more than 80% of all Imax Os.

The file system usually communicates with the storage media through the operating system Istroke O stack. Most advanced file systems are based on Linux and communicate through this stack, but the Linux stack adds overhead. Another way is for the file system to create its own Imax O channel to connect to the NVMe-based file system. From the point of view of the file system development process, direct communication with the drive is more difficult, but it provides the best opportunity for file system users to achieve maximum performance without over-compensation with expensive hardware.

File systems usually communicate with clients by using the standard NFS protocol, but NVMe has a network variant (NVMe-oF), and modern file systems should provide software support for parallelism and local NVMe-oF access to run on the client. NVMe-oF can also be used to interconnect various storage nodes, which makes the file system more accessible to directly connect the latency of storage.

In an all-NVMe file system architecture, metadata access is inherently fast, but metadata must be laid out efficiently to benefit from the low latency of NVMe. To optimize the performance of metadata, you need to span it across all nodes in the file system cluster so that there is no performance bottleneck for a single node.

How to make full use of NVMe

AI and high-speed loads can make better use of NVMe than other types of workloads. The challenge for these workloads is often that applications access storage through file systems, and traditional file systems do not optimize their IBO for NVMe-based drives. Faster node hardware and NVMe drives provide better performance, but the traditional file system architecture does not enable the hardware to reach its full potential.

To avoid this problem, you need to look for file systems that are written directly to the NVMe drive rather than through the operating system Imax O stack. Also look for file systems that allow clients to communicate across NVMe-oF and manage metadata in a way that does not affect performance.

Original author: George Crump Source: TechTarget

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.