> Vendors downplay the idiosyncrasies of specific SSD models by marketing their devices using four “headline” throughput metrics: sequential read, sequential write, random read, and random write.

For SOHO yes, where no serious database usage is expected. But server/datacenter SSDs are categorized: read-intensive, write-intensive and mixed-usage.

▲

wtallis 3 days ago | parent | next [-]

You're conflating two different things here: the performance metrics that marketing provides, and the product segments that marketing groups products into.

▲

p_ing 3 days ago | parent | prev | next [-]

Gamers also fall into the read/write number trap. When tested, that type of workload performs just about the same from PCIe 3.0 through 5.0 due to the 4KiB often random access. And in some cases, there was only a minor delta between PCIe 5.0 NVMe and SATA SSD.

https://www.youtube.com/watch?v=gl8wXT8F3W4

▲

antonkochubey 2 days ago | parent [-]

What games would load data in random 4KB chunks? Textures, sounds etc are in megabytes nowadays, 4K random reads are completely irrelevant.

	▲	p_ing 2 days ago \| parent [-]
		It doesn't matter how large the asset is, it matters what the method used to read the asset is. Not every application will read in a specific size, but 4KiB isn't uncommon.

▲

lmz 3 days ago | parent | prev [-]

Those categories are usually derived from another advertised number: Drive Writes Per Day.

As an example in this Micron product brief the Latency for the read-intensive vs mixed use product are the same: https://assets.micron.com/adobe/assets/urn:aaid:aem:e71d9e5e...

Of course the footnote says that latency is a median at QD=1 random 4K IO.

From the paper the PM9A3 which is 1 DWPD has better P99.9 write latency under load vs the 7450 Pro (3 DWPD mixed use).

	▲	bayindirh 3 days ago \| parent [-]
		The best way to spec a storage system for any use case is to give baseline numbers for the desired benchmark (plus its parameters), and let the vendors do their tests in house and spec the system out to you. If you can borrow systems, you can do it yourself, too. Otherwise, there are too many variables to calculate now. In the past it was easier. Now it's much more complicated.