What are the performance bottlenecks of traditional raid?

As we all know, traditional raid performs very poorly in Data Reconstruction. When a disk fails, it takes dozens of hours to recover the data in the faulty disk. Especially in the Process of Data Reconstruction, the performance of the Upper-layer file system will be greatly affected. In addition, when the application data is under pressure, the IO of Data Reconstruction and the application Io are staggered, leading to a vicious circle, resulting in poor performance in both Data Reconstruction and application Io performance.

 

This is the biggest challenge for large-capacity disks to traditional raid. Two years ago or earlier, many companies in the storage industry began to look for the next generation of disk RAID technology, the most likely and innovative of which is DDP (Dynamic Disk pool ), huawei in China calls this technology raid2.0. The biggest feature of DDP is that raid is built on a series of random chunks, rather than raid several disks. These chunks are the resource blocks in the disk, which can be 1 GB or larger. The concept of storage pool is common in DDP, which organizes these chunk resource blocks and makes them pooled. Obtain some chunks in this pool using certain algorithms, and then build raid based on these chunks. At present, many storage companies are developing such a new raid. At a glance, this technology seems very simple. In fact, many technical problems are hidden on the surface of the phenomenon, and it is not easy to solve.

 

This article does not want to talk too much about the technical details of DDP. I would like to discuss why the use of DDP can solve the problem of traditional disk raid. What is the performance bottleneck of disk raid?

 

To analyze the answer to this question, let's take a look at the disk raid data reconstruction process.

 

650) This. width = 650; "Title =" 1.jpg" src = "http://s3.51cto.com/wyfs02/M00/46/41/wKiom1Pwz3myIBSgAAEIUKigAsk986.jpg" alt = "wkiom1pwz3myibsgaaeiukigask986.jpg"/>

 

We can see that when a disk fails, other disks in the same disk group will have a lot of read operations to obtain the band data and then generate data on the faulty disk, finally, write the data to the spare disk. This is a complete data reconstruction process. All reads occur in the disk group of the faulty disk, and write operations only exist on spare. Therefore, we can easily find two performance bottlenecks:

• Spare disks are the performance bottleneck of write operations. If a single disk is written, the maximum write bandwidth of the SAS disk can reach about 130 Mb/s.

• The surviving disk in the disk group is a bottleneck in read operation performance. Because all data is read from a surviving disk in a disk group, the maximum bandwidth of the strip can only reach about 130 Mb/s.

 

Due to the above two major performance bottlenecks, the theoretical upper limit of Data Reconstruction bandwidth of traditional raid can only reach 130 Mb/s. To break through this theoretical upper limit, we can only break these two performance bottlenecks.

 

DDP can eliminate these two performance bottlenecks. Why? Because DDP is raid based on chunk, the chunk is obtained from a storage pool. Therefore, the layout of the chunk in raid on the disk is no longer as regular as that in the traditional raid disk group. In other words, DDP can achieve:

• The data write operations of spare chunk are no longer limited to one disk, but can be deployed in the entire storage pool. In this way, the performance bottleneck of writing a single disk will no longer exist.

• Refactored striped data read operations are no longer limited to a disk pool, but distributed across the storage pool. When the number of disks in the storage pool is large enough, the read performance of a single disk is no longer limited.

 

Through the above two points, DDP completely removes the two major performance bottlenecks of traditional raid, which can greatly improve the performance of Data Reconstruction. The read/write Io distribution on each disk after the DDP technology is adopted.

 

650) This. width = 650; "Title =" 2.jpg" src = "http://s3.51cto.com/wyfs02/M01/46/43/wKioL1Pw0ILw06v1AAExun48leY631.jpg" alt = "wkiol1pw0ilw06v1aaexun48ley631.jpg"/>

 

It can be seen that in the process of data reconstruction, there are read and write operations on each disk at the same time. Unlike the traditional raid, there are only a few blocks of inventory reading operations, and one disk has write operations. In addition, since the read/write operations are mixed, the disk may be jittery. Therefore, from the perspective of a single disk, both read and write operations can be greatly reduced. However, the overall data restructuring performance is enhanced as the number of disks involved in data restructuring increases. From this sentence, we can also understand that the application of DDP technology is limited. When the number of disks is not large, DDP technology cannot take advantage of its performance, will cause performance degradation. At present, SSD technology is quietly emerging, and there is no performance problem of read/write hybrid. Therefore, DDP technology is very suitable for SSD.

 

The biggest advantage after the introduction of DDP technology is that it can improve the Data Reconstruction performance of raid and increase the number of disks in the system. The Data Reconstruction performance can be linearly expanded. Traditional raid cannot achieve this.

 

650) This. width = 650; "Title =" 3.jpg" src = "http://s3.51cto.com/wyfs02/M00/46/43/wKioL1Pw0Enw9j_0AAEAPwH3F7Y474.jpg" alt = "wkiol1pw0enw9j_0aaeapwh3f7y474.jpg"/>

 

Today, flash storage continues to exist. In the future, many people estimate that the biggest reason for its existence is capacity (price). Therefore, they can build a hybrid Storage Architecture for the storage system. In large-capacity disk storage, we have reason to believe that the DDP technology will shine brightly and drive forward the traditional RAID technology.

 

This article is from the "Storage path" blog, please be sure to keep this source http://alanwu.blog.51cto.com/3652632/1541408