Three points to consider when comparing SSD performance
SSDs that store data in non-volatile semiconductor memory are superior in impact resistance, power efficiency, and read / write speed to HDDs that record and read information on disks coated with magnetic material, and the price has dropped in recent years. It became popular all at once. AnandTech, a technical news media, explains 'three features of SSD' that should be considered when conducting benchmark tests to compare SSD performance.
How We Test PCIe 4.0 Storage: The AnandTech 2021 SSD Benchmark Suite --Print View
◆ 1: Data recording method
SSD mainly consists of ' NAND flash memory ', ' memory controller ' and ' DRAM buffer '. NAND is the place to store data, the memory controller is the IC that controls the NAND memory, and DRAM is the cache memory that mediates between NAND and the memory controller.
In NAND flash memory, the circuit configuration required to store 1-bit information is called a cell. The more cells there are, the more NAND flash memory capacity, or SSD capacity, will increase. However, simply increasing the number of NAND flash memories will increase the area and price, so manufacturers have responded with multi-layered 3D NAND technology, miniaturization of manufacturing process rules, and MLC.
Of these, MLC is an abbreviation for 'multi-level cell', which is a recording method that increases the capacity of SSD by increasing the number of bits of information stored in one cell to two. As of 2021, for consumers, ' TLC ' that records 3-bit information in one cell has become the mainstream, and the cost per capacity is higher than that of 1-bit SLC or 2-bit MLC per cell. We have succeeded in keeping it low. Also, in recent years, 'QLC' that records 4-bit information per cell and 'PLC' that records 5-bit information have appeared, but the disadvantage is that the read / write speed slows down in exchange for a large capacity. there is.
TLC is also slower to read and write than SLC, and the number of reads and writes is also reduced. To solve this problem, most consumer SSDs with TLC NAND employ an ' SLC cache ' that records some of them in SLC. Faster read and write speeds on TLC by saving the data to be written to the SLC cache once, clearing the SLC cache as soon as the host I / O activity stops, and maximizing latency to prepare for the next write. It is a technology to make.
According to AnandTech, this SLC cache has a significant impact on benchmarks that write data to SSDs. Due to the limited size of the SLC cache, there are two types of performance measured by the benchmark: 'inside the SLC cache' and 'outside the SLC cache'.
Many SSDs in recent years have a variable SLC cache size, and as the remaining capacity of the SSD decreases, the available SLC cache size decreases. According to AnandTech, SSDs with QLC NAND use the SLC cache in addition to writing, so the SLC cache is more likely to be exhausted than SSDs with TLC NAND.
The following is Intel's NVMe (PCIe 3.0x4) connection / QLC NAND-equipped SSD '
In the range of general use, you rarely experience performance outside the SLC cache, but in a benchmark test that continuously writes tens to hundreds of GB of data, the cache will eventually overflow. It is quite possible that you will experience low performance outside the SLC cache.
◆ 2: SSD connection protocol
There are two main types of SSD connection protocols:
However, temperature rise is a concern for NVMe-connected SSDs. Therefore, until 2017, AnandTech tested each I / O pattern such as random access and sequential access for 3 minutes in each queue and continuous I / O for 18 minutes, but the exhaust heat of SSD is a problem every time. It was said that it was. Therefore, when conducting benchmark tests with SSDs with M.2 terminals and NVMe connections, AnandTech always wears a heat sink.
However, AnandTech said, 'While this kind of benchmarking test is useful for seeing how SSDs work under harsh conditions, it doesn't really matter how they work under normal use.' It states.
Also, even if the benchmark execution time is set, if a test that takes 10 seconds with a SATA-connected SSD is performed with an NVMe (PCIe 4.0) -connected SSD, 10 times more processing than a SATA-connected SSD will be required. .. As a result, AnandTech states that it puts limits on both 'running time' and 'the amount of data transferred in each benchmark test' to make benchmark tests more realistic and practical.
◆ 3: With or without DRAM bash
AnandTech points out that there are differences in the memory controller as well as the capacity, recording method, and connection protocol between the entry model and high-end model of SSD, especially the '
The flash translation layer (FTL) , which is one of the functions of the SSD memory controller, is responsible for allocating the block specified by the logical block address (LBA) to the physical block inside the SSD. The FTL address translation table that manages 1TB of NAND flash memory requires a little less than 1GB of capacity.
High-end model SSDs have a DRAM buffer memory to cache this address translation table. However, according to AnandTech, many entry-level SSDs do not have this DRAM buffer. With SSDs that do not have a DRAM buffer, read operations to the flash memory will increase during random access, and wear leveling and garbage collection will be difficult to manage. Therefore, if you perform a lot of write processing or the capacity is almost full, Performance tends to be significantly reduced. Therefore, there will be a difference in performance between high-end model SSDs and entry model SSDs.
Related Posts:
in Hardware, Posted by log1i_yk