anon (anon@anon.com) on 1/20/11 wrote:
---------------------------
>Caps can write out how much? A few hundred K of flash at most. And that's not necessarily
>enough to hold your trim data or mapping tables.
>
Enough. Actually for some devices the caps can write out the whole of the mapping table either by having enough capacitance or by having smaller tables. On others, they rely on check pointing of the mapping tables so that they only need to write out a small differential table on power failure. Cap in drives range from several small standard PCB caps to super caps. The actual energy to write flash is pretty low, it is the energy to erase that is high.


>And, "hundreds of us"? One hundred us is 10,000 iops and an Intel SSD can do what,
>40000 iops? Your high end capacitor backed etc etc drives may be able to process
>trim a bit faster, but they're doing what? Several times more IOPs than that too.
>
I'm aware of no flash based SSD that can do 40k iops. Most of the consumer level drives are in the couple hundred iops range (with 15K SAS drives being in the 50-150 iops range) and the enterprise drives generally cap out at a couple thousand iops. All for write. Reads are obviously higher.

Anyone who is quoting >10k random 4k wr iops to an flash based SSD is at best bending the truth and probably in marketing. AFAIK, no drive on the market can sustain >10k fullscan random 4k wr iops.


>You're just way off with all your numbers. I don't know where you come up with them.
>
If you can provide better numbers, be my guest.


>Um, no. A rewrite to the same LBA effectively *is* a trim of the old PBA. The old
>PBA is invalid at that point, the prior trim added nothing.
>
You aren't accounting for what the device is actually doing. A write to an unmapped LBA is faster than a write to a mapped LBA precisely because the drive doesn't have to do the work that the TRIM command does which is update the state of the unused/old PBA.