[Flash virtualization] software-defined server flash

Fan Jun (Frank fan) Sina Weibo: @ frankfan7 Twitter: @ franfan7

Challenges

When discussing the storage performance design on a virtualization platform with customers and storage vendors, we often face many challenges.

One storage plays multiple roles in the virtualization platform. For example, to meet storage capacity requirements (capacity), to meet application requirements (latency, iops, throughput), and to meet storage service requirements (snapshot, replication, backup ). Both capacity and performance must be satisfied. Therefore, a high-performance raid, high-speed hard disk, or a sufficient number of hard disks must be used to ensure performance. It is often found that the purchased hard disk is much larger than the actual capacity.

2. Is it designed based on the overall peak value or average value of the application? If the average value is used, the storage may be overwhelmed during the peak period, and the application requirements cannot be met directly. It can be designed according to the peak value, which will inevitably lead to excessive storage resources in most of the time and a large investment in the early stage.

III. The traditional storage architecture is centralized and shared. As the number or load of servers and virtual machines increases, the backend storage devices and storage network bottlenecks will sooner or later emerge. Is the upgrade solution flexible, simple, and easy to expand? What is the impact on online services?

Solution

Hybrid or full-flash Array

With the sharp decrease in flash memory prices in recent years, many manufacturers have begun to launch their own flash solutions. With regard to the Flash deployment method, some vendors add flash memory to the array to form a Hybrid Array together with traditional hard disks. Some simply launch all-flash arrays. Of course, the response speed of the match column is greatly improved, but it also has the following limitations:

First, IO requests from VMS to flash memory still need to be processed in many layers, which may be delayed due to queue congestion or storage network bandwidth.

2. From the perspective of the storage architecture itself, it is still a centralized shared architecture. In the case of performance bottlenecks that need to be upgraded, because of compatibility issues, it may not be possible to use the most cost-effective flash memory on the current market.

 

Server flash memory

Traditional server-side flash memory includes three forms: SAS-based interface, Sata-based interface and PCI-E-based. Server-side flash memory has the following features:

One is closest to the VM, so the latency in response to IO requests is the shortest. As shown in, IO requests must pass through six nodes to reach the array hard disk. Only two nodes can be processed by the server flash memory. Even if the storage network or array has a performance bottleneck, it does not affect the performance.

Flash memory technology is developing very fast, and the latest flash memory is often easier to use on the server.

Pernixdatafvp +Server flash memory

Adding flash on the server is the first step. How can we enable the virtualization platform to use flash? Pernixdata fvp combines the flash memory of all servers into a flash cluster, and can flexibly configure flash resources to datastore or VM. The concept of this solution is to separate the storage performance from the capacity, and the conventional array will continue to meet the capacity requirements, while the server-side flash memory will be used to take charge of the performance requirements.

The main notable features are:

1. application performance is greatly improved.

Most of the read/write IO requests sent by the VM are directly processed by flash memory, which greatly reduces the application response time.

Significant increase in second iops

The iops processing capability of flash memory is much faster than that of conventional disks in the array. The PCIe flash memory released by Intel is 250 kiops.

The cost of the three-Background array is reduced.

Since pernixdata uses the server-side flash memory to meet the storage performance requirements, the backend array only needs to meet the capacity. There is no need to use high-end arrays, expensive high-speed hard disks or expensive raid, which greatly reduces the cost of arrays.

Server-side Flash technology and pernixdata software-defined flash clusters provide a new generation of storage solutions with low cost, high performance, and high scalability in the virtualization storage design.

[Flash virtualization] software-defined server flash