SSDs used to be underestimated. Low reliability, high cost, lack of volume. Accordingly, hard disk drives and RAID arrays of varying degrees of complexity have been used more frequently to achieve higher speeds and increased reliability. Now the situation has changed noticeably, so flash memory is becoming more and more common. In personal computers, the main tip is already – SSD installation. But in the corporate segment, things are noticeably more complex and more flexible solutions are required.

Flash memory is being used more and more here, though. For 2024, SSDs are projected to account for over 42% of the server segment on a per volume basis. And will continue to grow. So, it’s a good time to talk about SSDs for the server.

Emphasis on performance

The disk subsystem, even built with RAID, is the bottle neck of the system. HDDs are slow, despite significant growth in volume, the speed of the drives has barely increased over the years. Accordingly, it becomes difficult to use for data caching, fast deployment of virtual machines and in systems with a predominance of random accesses. Using RAID to increase speed doesn’t give serious results either. Therefore, the use of SSDs in servers has become an inevitability.

Thanks to flash memory, owners of server hardware were finally able to breathe easy, because it became easier to deploy productive disk subsystems, which have become quite reliable, allow you to work with random access and are not a bottle neck in the platform. Modern SSDs have gradually started to surpass HDDs in everything.

HDDs are currently are only suitable as storage or in systems where the speed of the disk subsystem is not very important. But even in such servers prefer to place the operating system and the main applications on solid state drives, otherwise booting will take forever and launching any application will send the employee for a smoke break.

Of course, SSDs for server can noticeably speed up the system, but the price is still noticeably higher than hard disks, so the investment is still substantial. So, let’s talk about the problems of choice.

SSD for server
Computer and server SSDs

Which SSD for server to choose?

Solid-state drive – storagedevice based on flash memory. There are different types, such as, based on differentх methodsmethods of connecting the cells:

  • NOR is a two-dimensional matrix of conductors, withone link per intersection;

  • NAND is a two-dimensional matrix of conductors simply the transistor is replaced by a column of cells placed in series.

Modern “memory reservoirs” more often use the second option, because it is better in many ways:

  • high recording density;

  • memory erasure in blocks is done immediately, in NOR it is necessary to zero all bytes of the block first;

  • lower power consumption.

NAND in 2025 is more feasible. NOR-SSD is not visible on the market now, not even in the plans, so it is better to lead the narrative about the current direction.

So, it’s worth looking at the existing classic NAND types (we’ll look at 3D NAND later):

Flash memory type

SLC

MLC

TLC

QLC

Bits per cell

1

2

3

4

TBW

100k.

3 thou.

1 thou.

0.5 ths.

Reading time

25μs

50μs

75μs

110μs

Recording time

200-300μs

600-900μs

900-1350μs

1500 and more μs

Erase time

1.5-2ms

3ms

4.5ms

7ms

So, let’s decipher:

  • TDW – number of write/erase cycles of the memory cell, after the specified number of cycles thelink islikely to “die”;

  • all time values are for one cell. μs – microseconds, ms – milliseconds.

Okay, now it’s time to talk about each type separately.

SLC

Has one bit per cell, which reduces the load on the cell, consequently the link lives much longer. In addition, the single-level structure results in lower latency and higher read/write speeds.

Probably realized that the best type of NAND-memory is NAND-memory. Reliable, productive. Expectedly, the most expensive. The cost can exceed the price of “younger” brethren times.

For example, the Transcend 500TS64GSSD500 drive is built on SLC . The cost for 64 GB will be approximately 45,000 rubles. Not cheap. So it’s rare to see even in servers. There are more convenient alternatives that are practically as good as anything. Plus, the low recording density will make itself felt. Here you have to allocate each bit to a separate element, respectively, the physical size of the device will grow with the number of them. Accordingly, with large memory capacities combined with compactness, there is no need to speak.

MLC

Tiered. Has a higher recording density, because there are already 2 bits per cell, respectively, the cost is lower. But, had to sacrifice reliability. Currently one of the most common types of flash memory after 3D-NAND. For realization it was necessary to introduce additional 4 threshold voltages, but, as it turned out, the requirements to the condition of microchips increased, degradation of board properties more strongly affected the performance of components, so the number of rewrite cycles was significantly reduced.

In principle, the resource of modern MLC can be enough for several years of uninterrupted work, therefore, a good choice ofSSD for the server.

TLC

Subsequently, they decided to make a three-tier SSD to further increase the write density and reduce the final cost of the SSD. Unfortunately, this had a negative impact on speed and also severely reduced TDW.

It takes 8 levels of threshold voltage to realize such a structure. The demands on the chips increase even more, leading to more rapid degradation. So it was possible to make it cheaper, increase the recording density too, but the reliability was acceptable only for the consumer segment. As an SSD for a server, it won’t work.

QLC

NAND class devices of this type of memory have only been developed experimentally. The key difference is that there are 16 levels of charges per cell, that is, 4 bits per cell. Unfortunately, reliability became so poor that releasing such a solid state drive made no sense. But QLC implementation in3D-NAND is much more reliable, that’s why it was used there.

3D-NAND

The task of creating higher capacity fast and compact “reservoirs” with the release of previous types of “flash drives” has not lost its relevance. How to increase reliability and make the product cheaper at the same time?

Of course, science has found the answer to this question, after all, people can be very resourceful when the pressure is on. And the development gradually came to a multi-layered structure. Of course, there was still the challenge of fitting into the form factors, so you couldn’t increase the size of the drive. But making a “pancake” is very doable.

How to increase capacity?

  1. We add the number of bits for each memory link. Lose reliability, increase power consumption. The result won’t be good. And physically increasing the size is not an option either, because the market demands compactness.

  2. Techprocess. If the physical size of the link is minimized, more cells can be accommodated within a single board. Decided to go that route.

Reduced, but that was a bummer too. It turned out that it is physically impossible to reduce the size below 15nm, and the closer to the limit, the more exorbitant the number of defects grows. I had to look for another option to accomplish my goals.

It was decided to increase the number of elements not horizontally, but vertically, accordingly, within one chip began to fit more memory, and some of the problems that were characteristic of previous classes of memory, receded into the background. True, new ones have come in (increased likelihood of rejects, some cells malfunctioning), but rumor has it that they are being handled successfully. UPD: in 2019, the problems of broken cells have been successfully solved, the new drives are an order of magnitude better and more reliable.

For example,3D MLC NAND can have anywhere from 6,000 to 40,000 rewrite cycles and that’s not the limit. 3D TLC and QLC unfortunately have less TBW, but there are no exceptions here. Intel DC P4600 phenomenally marked a TVW equal to 11080. In fact, this SSD has 11080 write/erase cycles before link failure. The result is great. The cost is high, of course, but it’s for servers and data centers, you shouldn’t expect anything else. Expensive due to low latency per record. The average consumer doesn’t need it. Among consumer SSDs, 3,000-6,000 cycles is most common.

Intel SSD DC P4600

In general, a 3D NAND based SSD for a server is the most sensible option. Cheaper than the classic types, but with high reliability and excellent prospects.

Interfaces

So, uh. SSDS и HDD there are many interfaces. The most common ones:

  • SATA;

  • SAS;

  • NVMe.

Let’s break down each one, they have fundamental differences. Of course in terms of technical architecture there are almost no differences, but there is a serious difference in the approach to application. Therefore, I recommend to familiarize yourself with it, if, of course, the following is still unknown to you. At least the cheapest and slowest ones – SATA, SAS is 2 to 3 times faster. NVMe can outperform by a factor of ten.

SATA

This interface is designed for serial I/O requests. In fact, the bus does not allow more than one request in a row until a response is received. A clear sequence. It works poorly with asynchronous requests. For parallel work with a large number of references is practically not suitable. Severe loss of speed follows. Even flash memory can fail here, but it is the bus bandwidth that is the main problem. More than 6 Gbps is simply not available, and even then, only for streaming reads.

SAS

Serial Attached SCSI is also designed to handle serial requests, but already with two channelsand therefore random I/O requests are poorly realized, but noticeably better than the previous version, but only at the expense of 2 channels, no more. The bandwidth is up to 12 Gbps.

In mixed processes, it is better not to use even though the results will be noticeably more pleasingthan SATA. In addition, the price of a SAS SSDS is quite high, sometimes higher than the higher-performing NVMe-drives, which I will discuss next.

NVMe

A new logical interface designed specifically to handle mixed processes. Although it is serial, the PCI-E bus has a kind of point-to-point connection type, which allow you to break a request bit by bit and send it through several sequences at once and also receive a response.

In fact, Non-Volatile Memory Express is the fastest available, there are no limitations of previous interfaces, and any number of parallel accesses to the drive can be handled. Unfortunately, the price is noticeably higher than the previous two options.

But it is NVM Express that is the most promising interface at the moment. While the previous ones have speed limitations at the interface level, here the only limitation is PCIe . Now the bus is being fine-tuned to have even more bandwidth. Then the limitation will be the performance of the processor, RAM.

Intel Optane DC Persistent Memory

New memory from Intel. Placed in the same sockets as DDR4 memory. Non-volatile, so that when the device is de-energized, the memory contents are retained. The maximum capacity is currently 512GB, with higher capacity models in development. It’s not exactly a classic server SSD anymore, but in any case, other than the new Intel Xeon Scalable 2nd processors, nothing else supports these drives.

Intel Optane DC Persistent Memory

It has a DDR-T interface, in fact, the drive is installed instead of RAM, which eliminates the “layer” between the drive and the processor. Very useful in areas where you need to read and rewrite memory locations as quickly as possible. I am sure that in large analytics systems and in general when working with databases, there will be an application of this technology, but time will tell.

Conclusion

As you can see, there are many server SSDS. The choice depends on your objectives and the funds you are willing to invest in this endeavor. Good luck, I hope you manage to pick one componentswhich closesюneeds completely.

How useful was this post?

Click on a star to rate it!

Average rating 0 / 5. Vote count: 0

No votes so far! Be the first to rate this post.

If the materials from this website have been helpful, and you wish to support the blog, you can use the form at the following link: Donate to support the blog