On the relationship between storage iops and throughput throughput

Original link: https://community.emc.com/docs/DOC-29186?linkId=11675599

Introduced

ioPS and throughput throughput two parameters are the primary metrics for measuring storage performance. IOPS represents the number of storage transmissions per second, and the throughput throughput represents the total amount of data transferred per second. Both can represent the performance state of the storage in different situations, but the scenarios used vary. At the same time, there is a mutual connection between the two, this article on the IOPS and throughput throughput measurement of storage performance scenarios, describe the changes between the relationship and calculation method. Help readers better understand the performance analysis and planning of storage.

More information

The relationship between IOPS and throughput:

IOPS (io per Second) is used to calculate the number of transmissions per second in each node in the I/O stream (for each node in the IO stream, refer to the article:

Analysis of the relationship between the I/O process and the storage performance). Typically, the generalized IOPS refers to the number of I/OS processed by the server and storage System. However, because in the process of IO transmission, the packet will be divided into multiple blocks (block), to the storage array cache or disk processing, for the disk so that each block within the storage system is also considered an I/O, Cache-to-disk data processing within the storage system is also used as one of the metrics for metering in IOPS. The ioPS mentioned in this article refers to the generalized IOPS, which is the I/O unit that is initiated by the server and processed by the storage system.

ioPS are usually the most important metric for small I/O and the number of transmit I/O is large. For example, in a typical OLTP system, high IOPS means that the transaction of the database can be handled by the storage system.

Throughput throughput is used to calculate the total amount of data transferred per second in the I/O stream. This indicator, which is shown in most disk performance calculation tools, shows MB/s in the simplest copy of a Windows file. Typically, the throughput throughput computes only the portion of the data in the I/O package, and the data for the I/O header is ignored in the calculation of the throughput throughput. The generalized throughput throughput, also called "bandwidth", is used to measure the transmission channel in the I/O stream, such as 2/4/8gbps Fibre Channel, 60Mbps SCSI, and so on. But "bandwidth" includes the maximum of the total amount of traffic for all the data in the channel, while the throughput throughput protects only the actual data being transmitted, and there is a slight difference between the two.

Throughput throughput measurement is useful when minimizing time-consuming time for large I/O, especially when transferring certain data. Backing up the data is a typical example. In a backup job, we usually don't care how much I/O is handled by the storage system, but how much time it takes to complete the backup of the total data.

There is a linear variation between IOPS and throughput throughput, and the variable that determines their change is the size of each I/O. As you can see, when the I/O is relatively small, the time required for each I/OS to be transmitted is less, and the number of I/Os transmitted in a single unit of time is much higher.

Due to the processing of the packet header, the actual data transmitted in the total time is relatively low.

When I/O dimensions are large, as shown, the time to transmit each I/O increases and the iops count decreases. However, compared to the higher percentage of I/O channels used to transmit actual data, the throughput significantly increased.

We can use a simple formula to calculate the relationship between throughput and IOPS:

throughput MB/s = IOPS * KB per io/1024

Suppose you have a 10 10K SAS disk, each with a total IOPS of 1400 maximum IOPS. In theory, these disks handle different IO sizes, and the throughput throughput that can be achieved differs. In simple terms, the physical level IOPS and throughput which first reached the physical disk limit determine the performance threshold for this physical disk. The following calculation formula can see that the unit I/O size can multiply throughput, but fails to reach the theoretical "bandwidth" of 10 SAS disk 1gb/s (100mb/s bandwidth per disk). Obviously, because most applications do not have the same I/O, you will see that the throughput of the storage array is much smaller than the theoretical value provided by the vendor, because IOPS first reached the performance threshold, making the throughput no longer able to increase. Of course there are special applications, such as streaming media servers, the application side can use 2MB I/O size, the throughput utilization is obviously higher, the IOPS requirements are relatively low.

MB/s = 1400 * 64/1024 = 87.5 MB/s

MB/s = 1400 * 128/1024 = 175 MB/s

MB/s = 1400 * 256/1024 = + MB/s

In summary, when planning storage performance and dealing with storage performance issues, you need to look at the two parameters of ioPS and throughput throughput in a comprehensive view, this article summarizes the following points:

1. The throughput throughput of the performance tool statistics never reaches the theoretical "bandwidth" of the node in the actual I/O stream, because the performance tool does not count the I/O header information, but the actual data transfer volume.
2. Disk physical level IOPS and throughput which first reached the limit of the physical disk, determine the performance threshold of the physical disk, but determine which first performance threshold is the size of I/O.
3. Performance monitoring tool shows low IOPS or throughput lower than expected, do not directly consider the storage performance problems, to understand the application of I/O size, and then make subsequent judgments.
4. Storage performance Another important factor is the disk response time (Response times), which is based on the premise that storage can provide response time within acceptable access.

Reference

Analysis of the relationship between the I/O process and storage performance

On RAID write penalty (write penalty) and IOPS calculation

Description of different application storage IO types

Applied to

Storage performance

On the relationship between storage iops and throughput throughput