Must SSD be PCIe for enterprise-level servers?

I talked to users and vendors about how SSD (solid state drive) is used, I found that most people think that server SSDS must be ultra-high-performance SLC (single-layer unit) PCIe flash cards manufactured by manufacturers such as Fusion-IO, micron, Virident, or LSI. Although most people know that the mainstream SAS or SATA interface SSD is in the 2.5 inch format, they seem to think that these devices are only suitable for laptops or enthusiasts.

Of course, most sata ssds on the market do not meet the strict requirements of data centers, but there are several steps between low-cost SSDS and computing centers and non-cost PCIe cards. Most people worry that if SSD is configured in a disk array or server, these devices will exhaust service life and lose all data at a certain point in time.

I think part of the problem is that we are using devices in the data center instead of slowly consuming devices. The current peak decorative hard drive does not wear out in a foreseeable manner-they are randomly faulty-most storage systems regard the disk drive as a binary device, that is, either working properly, or you cannot work at all. Therefore, when a device experiences a front amplifier failure, it locks the connected RAID (Redundant Array of Independent Disks) controller to announce that the drive is broken and stops using the drive.

SSD write loss is not unpredictable. It is a relatively understandable process. The flash memory in SSD is well programmed and can be erased for a sufficient number of times until the array where the flash memory is located does not allow the flash controller to correct the error. This loss does not happen immediately on the entire SSD, but on a flash page and a flash page. In the end, This SSD does not have enough spare flash memory, and the SSD does not accept writing.

Most flash controllers carefully track how many times they overwrite on each page they manage. With the extension of the smart diagnosis system, the flash controller can report the write durability to the storage system. If I can know how many weeks or months the flash memory loss will cause me trouble, I only need to replace SSD when the loss rate reaches 80% or 90%.

The RAID controller can send a message saying that SSD 14 has reached the maximum durability, requiring a new or idle SSD and rebuilding the RAID set to the new storage. If the SSD is used to read the cache and does not contain unique data, it is easier to replace it.

Because the MLC (multi-tier unit) SSD (which may have only 5000 programming write-per-cycle durability) usually costs only one tenth of the cost of the eMLC or slc ssd. We may think of the mlc ssd as a consumable device in the data center. Although we may not be able to predict whether a mlc ssd can work for 18 months or 5 years, it is cheaper to buy three mlc ssds and replace them as needed than to buy one slc ssd.

Of course, this is not the way the data center buys the device-we purchase the device based on the forecast of the future peak needs, and then we need to multiply the coefficient we feel we need to ensure that the storage runs properly. Personally, I will multiply by π. Therefore, if a device may require 10 TB of storage, we need to ensure that the project budget can reach 30 TB, so that we will not worry about storage shortage in the future.

Of course, introducing the concept of consumable equipment in the data center may be more suitable for those who need to replace server SSDS close to service life. However, before we have excellent non-volatile memory in the future, it is also a way to avoid high capital expenditures and achieve high performance at the same time.

Some of you may worry that someone will recover sensitive company data from your discarded SSD. Although this concern makes sense, you can perform physical damage to the old lost SSD. If you can destroy the iPad as a waste, you can also do this for micron P400 RealSSD.