IOPS & Latency: What are the performance determinants

Latency is the concept of processing speed of an I / O request of the system. This concept is important because a storage system, although only having a capacity of 1000 IOPS with an average latency processing time of 10ms, can still outperform a system with 5000 IOPS but a latency of 50ms. Especially for "sensitive" applications with latency, such as Database Services.

Take a practical example, in a supermarket, if the cashier (hard drive) serves customers (I / O) with latency time is 10ms. So, it can be understood that this cashier serves 100 customers / 1 second. However, what if there are 100 guests coming at the same time within 10ms? Customers will have to wait. And depending on customer needs (size I / O), latency may vary, 15ms or even 20ms.

The following illustration will show you why an increase in latency in IOPS is higher. The truth is that the storage system will look at the queue (queu) and order sequential processing for I / O, leading to a longer queue, the latency will be higher, depending on the performance of the application. Whether we can accept this or not.

In order to compare storage system efficiency, platform and application environment factors need to be the same - this is difficult, because enterprise systems need to run multi-workloads. In some cases, it is good to process / transfer large amounts of data (high throughput), but when it needs to process large numbers of small I / Os quickly (requires IOPS), it is unlikely and vice versa. again. Now the size of I / O, the length of the queue (queu depth) and the level of parallel processing ... all have an effect on performance.
IOPS - Perhaps the system using the current HDD or SSD is too high, when standing alone, this number becomes useless. And intangible, it becomes a term for manufacturers to market their devices, businesses should not be included in it as a measure of decision on storage system performance.
Instead of asking: "How many IOPS is the system with? ” We should ask: "What is the application processing time?Latency should be considered the most useful parameter, because it directly affects the performance of the system, which is the key factor that should be based on IOPS output and throughput. This means that minimizing latency will help improve overall system performance