The use of solid state technology in the form of NAND flash for storage systems changes the way we need to evaluate storage. While it brings power, space, and reliability advantages, the main reason for using solid state is performance – it accelerates applications.
Still, storage vendors often characterize their flash systems in ways more fit for spinning disk. The numbers usually quoted for storage systems are:
- IOPS – The number of I/Os per second that a storage system can do.
- Bandwidth – The measure of throughput for sustained data transfer as an MBps or GBps number. Bandwidth measures are important when dealing with a high volume of data. The data transfer rate is dependent on the size of the data transferred (block size) and the overhead processing between each block. Infrastructure options — including the fabric and the network reliability to ensure transfer completions without needed retries — are also considerations.
- Latency – This represents the time required for an I/O operation to complete.
These need to be considered differently with an all-flash storage system. First, IOPS is an aggregate number for a system but does not indicate what benefit an application or workload will achieve from a particular system. As an aggregate number, it can be deceiving based on the size or scale of a system. The maximum number needs to be placed in context of the overall workload supported by the entire storage system.
Latency is the more important measure for solid state. Performance measurements go back to mainframe storage systems, when disk rotational latency and the ability to queue I/Os were individually measured. Response time was the important measure because it included latency and queue time. With a storage system design based on solid state technology, queuing is not the major factor as it is with electro-mechanical devices such as disk drives.
Many people are even mixing the terms response time and latency, and some vendors use them interchangeably. Service time is latency and the data transfer time. Latency with disk drives is seek time plus rotational latency before the data transfer completes. However, vendors do not usually quote service time and use the term latency because it looks better comparatively than the latency of spinning disk. Data transfer is limited by the interface and network connections so that is less of a product advantage to highlight.
Given that vendors quote latency and response time interchangeably without referring to service time, those two values must be used for comparative evaluations. Important factors are how fast an I/O completes, and that it is a predictable amount. High variations in I/O completion create problems in management in addition to the effect on business.
For all-flash arrays, latency or response times are the important measures. IOPS needs to be discounted. Because it is an aggregate measure, IOPS will not give you a good understanding of acceleration and the value that brings. How fast the I/O can be completed is the most important factor in the first level consideration for solid state storage.
The Evaluator Group has additional guidance on how to measure solid state storage performance.
(Randy Kerns is Senior Strategist at Evaluator Group, an IT analyst firm).