[lustre-discuss] File migration from one OST to another for Lustre 2.5
andreas.dilger at intel.com
Wed Jun 24 00:52:28 PDT 2015
On 2015/06/24, 1:39 PM, "Tung-Han Hsieh" <thhsieh at twcp1.phys.ntu.edu.tw>
>I have a question about data file migration from one OST to anther
>for Lustre 2.5.
>Suppose that OST0000 is going to be removed. In the past (I mean,
>Lustre-1.8.7), I can do the migration smoothly via the following
>1. In MDT server, stop writing new files to OST0000:
> lctl set_param osc.foo-OST0000-osc.active=0
>2. Find out all files located in OST0000 (suppose that the lustre
> filesystem is mounted at /home):
> lfs find --obd foo-OST0000_UUID /home > OST0000.txt
>3. For each file listed in "OST0000.txt", I did:
> cp -a file file.tmp
> mv file.tmp file
> So each file in OST0000 will be replaced by a new copy created
> in other OSTs. If I login into the OST0000 server and use df
> to check, I can see that the disk usage of the OST0000 partition
> is fewer and fewer. Eventually it will go to nearly empty.
>4. Finally, remove OST0000 permanently:
> lctl conf_param osc.foo-OST0000-osc.active=0
>But for Lustre 2.5, when I repeated the above procedures 1 to 3
>(I haven't done step 4), the disk usage of OST0000 still remains
>the same, while the disk usage of the other OSTs keep increasing.
Correct. This is a change that has been introduced since Lustre 2.4 in the
way that OST objects are destroyed. The MDS is now in charge of OST
object destroys, so if its OST connection is disabled then the OST objects
will not be destroyed until the MDS connects to the OST again. Please see
LU-5931 for more details.
>Then I interrupt step 3, and check step 2 again, I did see that
>many files were disappeared from OST0000. So the migration works.
>But since the disk usage of OST0000 did not change at all, it means
>that the migrated files still left the "bodies" in OST0000. They
>become junks in OST0000.
Which is fine if you are removing OST0000 from the filesystem permanently.
If you aren't removing it permanently, it will be fixed up when the MDS
reconnects to it again.
>Is it the normal behavior of Lustre 2.5?
Yes, though it could be improved.
>If I want it to behave exactly the same as Lustre 1.8.7, how should I do ?
Sorry, that isn't possible today.
Lustre Software Architect
Intel High Performance Data Division
More information about the lustre-discuss