More frequently IT departments are looking at Flash as an efficient and cost-effective method of increasing transaction rates and supporting more concurrent users without having to add servers or alter their applications. For executives and administrators unfamiliar with the differences between hard disk drives and Flash, the competing claims and specifications of vendors can easily be confusing, misleading, and sometimes daunting. In this article, we will explain how Flash media is written and erased, and how these processes differ from the equivalent processes for hard disk drives. This will help readers to understand why initial, or burst, performance rates for Flash media may differ from the sustained performance rates that users can expect to see in steady-state operation. By demystifying the technology, users can make more informed Flash system comparisons and better procurement choices to meet their IT needs.
To begin, let’s remember that a hard disk drive is a magneto-mechanical device. If you want to write 01001011, you issue the command to do so. A moving “head”, floating above a spinning platter, alters magnetic poles on the platter – 1s become 0s or 0s become 1s and the data is stored on the drive. To change the data values, you simply overwrite the original data with the desired vales. The head again alters the magnetic poles of the spinning platter and new bit values are stored in place of the original ones.
Follow the sequence…
01001011 --- Original desired data impression
01100110 --- Altered impression
The process is simple and relatively straightforward, although there is time-consuming latency involved as the drive spindle and head move mechanically to the correct location on the drive. Typically, the rotational latency of a hard drive is between two and five milliseconds. An individual disk can sustain a maximum of around 200 random I/Os per second (IOPS). Although the rewrite process is a single-step, performance is orders of magnitude slower than with a Flash system.
In the case of Flash, data is stored on solid-state memory chips and there are no magneto-mechanical parts. With Flash, changing data values is a one-way ticket. At the bit level, a clean Flash page starts with all 1s in its storage cells. Writing data to this page either leaves each cell at the value 1 or flips each cell to the value 0. A command to change the data to a combination of 0s and 1s sends an electrical pulse to the Flash chip in order to switch the desired 1s to 0s. Because there is no rotational latency, this happens rapidly. Flash latency in an
In a new Flash system, all the chips are in the erased (all 1s) state and so can accept write changes immediately and rapidly. Measuring this performance gives an initial optimal performance reading known as a “burst” rate. Many vendors publish this rate in their performance specifications. However, this is not the rate that is likely to be sustained once the data on the Flash chips needs to be rewritten.
Unlike hard disk drives, where polarity is changed - albeit slowly - Flash cannot change individual 0s directly to 1s. The chip needs to be erased, or re-initialized to the all 1s state, before it can be rewritten with the desired 0 and 1 combination.
Again, let’s follow a sequence…
11111111 --- Erased impression
01001011 --- Original desired data impression
11111111 --- Erased impression
01100110 --- Altered impression
Note the extra steps that took place to get to an initial impression and then the altered impression in Flash.
Sustained performance in Flash systems can be increased by way of two mechanisms: increasing presented capacity and increasing chip count. By reserving a sufficiently large pool of Flash capacity that is pre-erased, there is a larger available pool for immediate writes. With a ready pool of erased and writable Flash available at all times, the system does not suffer an erasure delay so sustained performance increases. When choosing a Flash system it is important for buyers to understand whether the system has an adequate reserved erasure pool, and if the size of this pool can be modified. When comparing price/capacity of competing systems, usable space or presented capacity is more meaningful than raw capacity comparisons. However, comparing price-performance metrics for sustained performance is really a more useful metric than price-capacity in making a sound economic comparison.
Sustained performance is also improved with a higher chip count. A larger number of chips allow the Flash system to sustain high burst rates for longer and deliver a higher sustained level of performance. This is because there are more chips to manage the read, write, and erase cycles. With a sufficient volume of chips, erase operations do not occur serially. Operations can instead be executed in parallel, with reads, writes, and erasures all occurring at the same time and at very high performance rates. The minor compromise with having the higher chip count necessary to sustain high Flash performance is that extra chips require more real-estate, so form factors for cards with higher sustained performance will generally be larger than for cards with lesser sustained performance, for instance a full-size versus a half-size PCIe card.
While Flash undoubtedly offers the potential for applications to make a quantum leap in performance and productivity, I/O-intensive applications need Flash systems that can sustain performance in order to handle the ongoing needs of servers handling parallel queries, large numbers of concurrent users, and the I/O for their various nodes. Before buying a Flash system, the IT department needs to understand exactly how their Flash system will operate under ongoing and peak conditions in their environment. As they evaluate systems, IT should stress the system with large I/O volumes and watch how the performance peaks and then levels after at least a full day’s use. Experienced vendors who stand behind their products will be willing and able to verify sustained performance metrics so that their customers realize optimal value from their Flash SSD investment.
In summary, there are distinct functional differences in read and write behaviors between a solid state disk and a hard disk drive. There are also major differences in what the user of an SSD implementation can expect in terms of performance out of a specific solution based upon multiple factors such as type of Flash utilized, architectural choices, wear leveling algorithms, and design implementations. Burst rate versus sustained rate must be clearly understood in order for IT managers to have realistic performance expectations via the deployment of SSD into their architectures. And in the end for those IT managers to have the best possible experience with what has shaped up to be the most disruptive, yet promising technology break through in storage hardware’s recent history.
