Linus Torvalds (torvalds@linux-foundation.org) on 1/21/11 wrote:
---------------------------
>Ricardo B (ricardo.b@xxxxx.xx) on 1/21/11 wrote:
>>
>>Can *anyone* point out those mythical SSDs
>>
>>- for which TRIM reduces performance, either in the short or long term.
>
>Umm. It's more like the reverse.
>
>The _common_ case is that trim reduces performance, because
>it makes file delete slower.
>
>In fact, I can pretty much guarantee you that it does that
>on just about every ssd that supports trim out there.
>
>And on some systems, trim just makes them unusable, because
>the ssd is so bad at it.
>
>Of course, part of it is the actual trim patterns, but
>see for example
>
>http://people.redhat.com/jmoyer/discard/ext4_batched_discard/ext4_discard.html

Wow, that's ugly. The first 3 seem alright though.

>just to see some numbers, since people seem to be in denial.
>
>Yes, those "1000%" things are real. See how the "good"
>devices only slow down by 15-20%, while the bad device gets
>ten times slower when the filesystem sends discards for
>blocks as it deletes them.
>
>Yes, yes, so people are trying to come up with better
>ways, where you discard in batches, and where you just
>discard entirely off-line. As mentioned, my personal
>opinion is that on-line trim is just stupid, and doing
>it offline (when the system is quieter - so not necessarily
>_totally_ offline, but basically not during real ops)
>is the only sane approach.

Can you clarify what you mean by 'on-line' and 'off-line'?

My intuition is that if you have a period where the SSD is quiesced (say 200us where you aren't accessing the drive), that would be a fine time to do a little TRIMming.

OTOH, if you are being slammed with a ton of IOPs, then I can easily see how TRIMming could be bad as you contend with real drive activity.

Certainly, I've heard reports of the GC in the drive causing incredibly high latencies (for bad GC algorithms) due to blocking incoming I/O.

So how much time does 'off-line' mean? Do you really mean that the drive has been taken out of the FS entirely?

>But dammit, can you guys shut the f*ck up already? Admit
>that you were wrong, and trim isn't fast, and isn't a
>simple thing. Because IO patterns matter, and without ESP
>you can never know when it's a good idea to TRIM.

Do you mean ESP as in "magic crystal ball", or is this a technical term I'm not familiar with? : )

David