Why are NVMe SSDs Faster? Here's the Secret
Toshiba RC100 is a special NVMe SSDs. The first impression is "small" size. On account of Toshiba BiCS3 high-density 3D flash memory and MCP multi-chip fusion packaging technology, the small size SSD contains a lot of energy.
The NVMe SSD is the best among SSDS. Why does SSD with NVMe protocol have better performance than SATA? This is not just due to the bandwidth advantage of the PCIe channel.
Surface performance and practical impact of SSD
Just as CPU operations have integer and floating point, computers have sequential and random reading and writing of SSDs. Although we can measure the theoretical capability of SSDs in both access modes by testing software, we cannot tell the direct impact of such data on computer speed.
The bandwidth advantage of the PCIe channel is mainly reflected in the sequential reading and writing speed, which is the Seq (sequence) line of the AS SSD Benchmark in the figure below. However, sequential reading and writing usually only occurs when files are copied between two hard disks, which is a less common use for system disks of home computers. Similarly, although the system disk IO is mainly based on random reading and writing, it is not only a 4KB granularity, nor is it a separate reading and writing.
Finally, these figures provided by AS SSD Benchmark are mainly to confirm theoretical parameters, which have no strong practical reference value.
Key factors affecting SSD performance:
Sequential reading and writing bandwidth and 4K random reading and writing IOPS are easy to measure, but not very useful. The key to hard disk performance is IO latency. IO latency is the amount of time a reading and writing operation takes from start to finish. The lower the latency, the faster the hard disk responds to the host's reading and writing requests and the better the user experience.
Some friends may ask can the delay not be directly replaced by the IOPS indicator? The answer is no. Unless the IO comes one after the other, the delay can be equal to 1/IOPS. This situation is so ideal that it does not exist in reality.
Whether it's a mechanical hard drive or a solid state drive, one IO has not been completed and another IO reading and writing request has arrived. Mechanical hard disk can only have one magnetic head to perform reading and writing operations at the same time, so it can only execute the next IO request after the previous IO is completed and the delay of subsequent IO requests will inevitably increase.
Although from AHCI protocols have been introduced NCQ features that has allowed the individual IO instructions in a queue at the right time "to jump the queue" and adjust execution order that the hard disk heads have the opportunity to reduce the frequency of seek, but no matter how to optimize, IO or had been executed one by one, the increase of the queue depth seems to increase the number to finish the reading and writing (IOPS) per second, but actually IO delay is increased.
SSDs use flash memory without moving parts to store data, eliminating magnetic head seek time in mechanical hard drives and can reading and writing concurrent data through multi-channel and CE interleaving technologies. But the nature of disk performance determined by IO latency has not changed.
The PCMark 8 storage performance test measures the impact of a hard drive on performance by using a series of hard-drive activity scripts in a real-world application. Behind the performance of the PCMark 8 storage performance test is still access latency. Average access time or average access latency can be seen when we save the PCMark 8 storage test scores in XML format.
IO latency is a difficult metric to quantify and hard disks can perform differently in different applications. To reduce the delay, in addition to flash interface, master design, NVMe protocol is one of the "shortcut". NVMe protocol for flash memory optimization has made a series of improvements for low latency, effectively reducing latency overhead.
Thanks to the help of NVMe protocol, Toshiba RC100 SSD can easily cope with many complex applications, effectively complete data reading and writing work and improve the computer experience with the advantage of low latency.