anon (anon@anon.com) on 1/20/11 wrote:
---------------------------
>SRAM/DRAM? Battery backed? I'd be surprised if that was the case, on a commodity SSD.
>
Almost all the SSD controllers on the market utilize off chip DRAM for tables (and some even for caching data). The only exception I know of is the sandforce controller but it tends to have bad performance when it has to use RT GC precisely because it is a sliding window/table small onchip map. We are talking about the full on consumer drives. Almost all of them utilize a single DRAM chip primarily to store mapping tables. The DRAM chips used vary in size from 32MB to 256MB. Just look at board photos for any of the Marvell/Indilix/Jmicron/Intel based SSDs which I'm sure you would agree fit into the commodity SSD range (since with the Sandforce controller, they are the only controllers used).

Several of the drives utilize capacitors as well. And certainly all the enterprise drives do. Not battery backed, but capacitive power in case of a power loss to write the mapping tables. There are only two other options: don't worry about power loss or do full device scan to rebuild tables on power on.


>Anyway, right. So when you say "it's free, it just has to update this table", you
>mean that an IO has to go down the wire, and the device has to do some processing
>and potentially some banging on NAND (or PCM or whatever) or even this SRAM, and then send the completion back up.
>
>That's only "free" for some restricted definitions of "free".
>
Its on the order of hundreds of nS and device latencies are on the order of hundreds of uS. I consider < 0.1% to be free.

>And even on those, TRIM will be a total loss if the filesystem rewrites the same
>LBA before the old one is garbage collected.
>
No, not a total loss. For a write to an unmapped LBA, you don't have to do a map and mark to free/dirty. So it saves there. If you are doing block erase and write, it shouldn't have any effect and if you are doing random IO, the device should have plenty of free time.

In a properly implemented SSD, the overhead for doing a TRIM really is minimal. And the occupancy/latency on the SATA/SAS interface should also be in the noise.