1 |
On 04/09/14 20:07, Rich Freeman wrote: |
2 |
> On Thu, Sep 4, 2014 at 2:26 PM, Сергей <protserovsd@×××××.com> wrote: |
3 |
>> You need to run Fstrim if you mounted your partition WITHOUT "discard" |
4 |
>> option and did lots of changes. For example, if you installed your |
5 |
>> system without "discard", do fstrim and then add "discard" to |
6 |
>> /etc/fstab. |
7 |
>> |
8 |
> Just a note that depending on the SSD model, discard can have a |
9 |
> substantial performance penalty with negligible benefit compared to |
10 |
> just sticking fstrim in your crontab. |
11 |
+1 |
12 |
also for lvm remember to add in lvm.conf |
13 |
issue_discards = 1 |
14 |
|
15 |
> |
16 |
> In theory the ssd should just handle discard by making a note of what |
17 |
> is trimmed and utlizing this information when needed. In practice |
18 |
> many ssds handle a trim by dropping whatever they're doing and don't a |
19 |
> copy/delete cycle if only part of a block is trimmed, which defeats |
20 |
> the whole point of trimming in the first place. FStrim has the |
21 |
> advantage of being more asynchronous and possibly being able to |
22 |
> consolidate trims over a longer time-period, which could improve |
23 |
> performance if the ssd isn't smart about it. |
24 |
i understand that part of the reason it blocks so hard when run and |
25 |
hasn't been run is the ssd defrags/consolidates the used blocks too. it |
26 |
would be nice to know for sure what it's doing, or from kernel-space |
27 |
tell the ssd what is most likely to be changing and what can be |
28 |
consolidated happily. |
29 |
> |
30 |
> Is there a really good place to go for SSD reviews/etc that actually |
31 |
> takes this sort of thing into account? After getting an SSD it became |
32 |
> apparently that they vary widely in terms of quality. Heck, I can't |
33 |
> even tell you what the erase cycle count is from the SMART info, while |
34 |
> other models seems to provide all kinds of useful info. |
35 |
+1 for this as other factors such as the erase blocksize should be taken |
36 |
into account. i.e. the larger the erase blocksize the more need there is |
37 |
for fstrim in the first place, but also the filesystem/partition |
38 |
alignment becomes a magical dark art. |
39 |
> |
40 |
> -- |
41 |
> Rich |
42 |
> |