Shulou(Shulou.com)06/03 Report--

Author: Frank Fan Sina Weibo: @ frankfan7 Wechat: GetToCloud

In the sixth Latnecy of practical storage design, we introduce the reasons why Latency is too high and some suggestions. This paper focuses on the impact of Block Size on performance, especially latency.

What is Block Size?

The word Block Size has different meanings in different contexts. In this article, it refers to the size of the data unit that an application running in OS sends an IO read or write request. Whether you use traditional storage architecture, or converged or super-converged architecture, IOBlock Size has an impact on performance.

As shown in the following figure, an IO request, such as moving a brick, the size of the brick has an impact on the load of many components in the entire storage architecture. For example, HBA, storage switch, storage system CPU and so on, of course, the storage media itself has different processing capacity for different bricks.

The above picture is taken from the blog vmpete.com.

Why does Block Size have such a big impact on storage performance?

Impact on Throughput:

Throught = IOPsx Block size

The same 10IOPs 256k blocks needs to transmit 64 times more data than 4K Blocks! Then it will naturally affect the bandwidth of the storage network and the CPU consumption of the storage controller.

Influence on delay Latency

The above picture is taken from vmpete.com. This shows the impact of different Block Size on Lantency after using the storage performance analysis software PernixData Architect. It can be seen that when Block Size is greater than 64K, it has the greatest impact on Lantency.

How to deal with the challenge brought by big data block IO?

Previously, there were no tools or simple methods to understand the Block Size generated by the application. Moreover, in the conventional storage design, the Block Size generated by the application is not taken into account, and the design decision is often made according to some assumptions. Even if flash memory is used in the storage array, it cannot be reduced. Big data block IO load on HBA, storage network and storage controller. And we find that the performance of conventional SSD hard disk processing big data block IO is not ideal.

Take advantage of the real case of IO BlockSize:

Case 1:

A customer found that the application performance decreased, and after debugging, he found that big data block IO had a great impact on Latency. PernixData FVP software and server-side NVMe PCI flash memory are used to accelerate storage performance without any changes to background storage. It greatly reduces the influence of big data block IO on Latency. Note that this user uses NVMe PCI flash memory instead of regular SSD Drive. Because the conventional SSD hard disk processing big data block IO should not be good.

Case 2:

A customer found that the performance of many applications degraded. After debugging, it is found that a SQL virtual machine produces a lot of big data block IO, and the problem is solved after optimizing the SQL.

Case 3:

A cloud service provider sets up Disk IO Limit in all virtual machines to prevent some virtual machines from overusing storage resources in a shared environment. Mclock diskscheduler has been changed since vSphere 5.5. for IO whose block size is greater than 32K, from the point of view of Disk IO limit, the IO is no longer considered to be an IO. For example, 256k blockSize IO would be considered eight IO.

Getting an overview of the application of Block Size will be of great help to better set up Disk IOLimit.

Reference:

Viewing the impact of block sizes with PernixData ArchitectLimit a VM from an IOps perspective

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.