Can I create a traditional raid on an SSD?

When asked whether traditional raid can be built on SSD, the first response should be "yes, but trim command support is required ". On the Internet, we can also see that many people use SSD to create raid, but basically it is configured as RAID 0 or raid 1. Few people will build RAID5. Why?

The answer is very simple. If you configure parity-raid like RAID 5 on the basis of SSD, can the integrity of SSD be well realized? More importantly, parity-raid will affect the life of SSD. In a word, traditional raid kills SSD. Therefore, traditional raid cannot be directly configured on SSD, not only because of the support of trim commands.

Before analyzing the problem of traditional raid on SSD, Let's first look at what value raid can bring to the entire SSD? Tom's Hardware configures a RAID 0 on the basis of SSD, and compares the performance of raid0 and a single SSD. It can be seen that when the two SSDs are configured as RAID 0, the performance can basically double whether it is sequential read or write. Therefore, you can increase the overall throughput of the system after configuring raid0. The sequential read/write performance in raid0 mode is compared as follows:

 

650) This. width = 650; "Title =" 1.jpg" src = "http://s3.51cto.com/wyfs02/M01/30/9B/wKioL1OpFkLBgv_9AAEOA-7MusU520.jpg" alt = "wKioL1OpFkLBgv_9AAEOA-7MusU520.jpg"/>

 

The second concern is the random read/write capability. One of the major advantages of SSD is its high random read/write capability and high iops. Therefore, after being configured as raid, can I double the iops capability? The answer is not necessarily. When the queue depth is very small, there is no gap between the random read/write capability in raid0 mode and the performance of a single disk. When the queue depth is 1, the random read/write performance in raid0 mode is compared as follows:

 

650) This. width = 650; "Title =" 2.jpg" src = "http://s3.51cto.com/wyfs02/M01/30/9B/wKiom1OpFpShu0GuAAEvgkwKA50413.jpg" alt = "wkiom1opfpshu0guaaevgkwka50413.jpg"/>

 

The random read/write capability of raid0 can be improved only when the queue depth reaches a certain level. When queue_depth reaches 64, the random read/write performance is as follows, which can basically be twice the performance of a single disk:

 

650) This. width = 650; "Title =" 3.jpg" src = "http://s3.51cto.com/wyfs02/M01/30/9B/wKioL1OpFouwlrF4AAEw-pxIiF4409.jpg" alt = "wKioL1OpFouwlrF4AAEw-pxIiF4409.jpg"/>

 

The third concern is latency. In databases and other applications, I/O latency indicators are particularly important. Can latency indicators be reduced after raid is created? In principle, raid can only prolong the I/O latency and cannot reduce the I/O latency. The test results of Tom's hardware are as follows:

 

650) This. width = 650; "Title =" 4.jpg" src = "http://s3.51cto.com/wyfs02/M00/30/9C/wKiom1OpFt2hf9flAAEsHmlFQN0014.jpg" alt = "wkiom1opft2hf9flaaeshmlfqn0014.jpg"/>

 

We can see that the latency increases after raid0 is introduced. Therefore, raid can improve the overall throughput of the system. When the queue depth is large, it can improve the random I/O capability of the system. If the latency is not improved, it can only increase the latency; if you configure it as parity-raid, you can also improve data reliability.

In many applications, raid is required to support high Io throughput, iops, data reliability, and large capacity. However, as mentioned above, traditional parity-raid cannot work well on SSD media and can only accelerate the wear of SSD. The main problems of traditional raid are as follows:

1. Raid partial-stripe write problems. In parity-raid such as RAID5, if partial-stripe is written, you must first read the unupdated data in the Strip and then merge it with the newly written data, finally, calculate the check value and write the new data and the check value to the disk. The entire partial-stripe write process is a "Read-Modify-write" operation ". From the performance perspective, the write process of partial-stripe seriously affects the performance and increases the IO latency. More importantly, partial-stripe writes frequently to update the parity data. Because SSD adopts the out-of-place data update method, this frequent parity data update will increase the SSD write amplification factor and affect the service life of SSD.

2. Data restructuring. Traditional raid Data Reconstruction is for disks. That is to say, when a disk in raid fails, raid removes the disk from the raid group and finds an idle spare disk for Data Reconstruction. Disk-level data restructures the SSD data from the beginning to the end, that is, the spare SSD will be written from the beginning to the end. In SSD, FTL is used for page/block ing. If all page pages in SSD are exhausted, the SSD will be forced to start the garbage collection, making the SSD performance the worst.

3. Data Synchronization problems. This problem is similar to the Data Reconstruction problem.

4. the SSD disk is faulty at the same time. Parity-raid evenly distributes the data in the Strip to all disks, so that the IO on all disks is equal. The lifetime of an SSD is basically the same. Therefore, in raid mode, the probability of simultaneous failure of SSD in a short period of time is very high.

The underlying technology of an SSD disk is completely different from that of a disk. Traditional raid cannot work well with SSD disks, which seriously affects the overall performance and SSD Service Life. Therefore, creating raid on an SSD is not that easy.

 

 

This article is from the "Storage path" blog, please be sure to keep this source http://alanwu.blog.51cto.com/3652632/1430288