Thanks Greg :)
For the OSDs, I understand, on the other hand for intermediate
abstractions like hosts, racks and rooms, do you agree that it should
currently be possible to change the IDs (always under the "one change at
a time, I promise mom" rule)?
Clearly, a good amount of shuffling should be expected as a consequence.
Basically I was inquiring whether changing the id of a single host would
shuffle the entirety (or a relatively big chunk) of the cluster data, or
if the shuffling was limited to a direct proportion of the item's weight.
I just --test-ed with crushtool. I changed an host's id, and testing the
two maps with :
crushtool -i crush.map --test --show-statistics --rule 0 --num-rep 3
--min-x 1 --max-x $N --show-mappings
(with $N varying from as little as 32 to "big numbers"TM) shows that
nearly the 50% of the mappings changed, in a 10 hosts cluster.
Thanks All :)
Le 02/11/2015 16:14, Gregory Farnum a écrit :
Regardless of what the crush tool does, I wouldn't muck around with the
IDs of the OSDs. The rest of Celh will probably not handle it well if
the crush IDs don't match the OSD numbers.
-Greg
On Monday, November 2, 2015, Loris Cuoghi <lc@xxxxxxxxxxxxxxxxx
<mailto:lc@xxxxxxxxxxxxxxxxx>> wrote:
Le 02/11/2015 12:47, Wido den Hollander a écrit :
On 02-11-15 12:30, Loris Cuoghi wrote:
Hi All,
We're currently on version 0.94.5 with three monitors and 75
OSDs.
I've peeked at the decompiled CRUSH map, and I see that all
ids are
commented with '# Here be dragons!', or more literally : '#
do not
change unnecessarily'.
Now, what would happen if an incautious user would happen to
put his
chubby fingers on this ids, totally disregarding the warning
at the
entrance of the cave, and change one of them?
Data shuffle? (Relative to the allocation of PGs for the
OSD/host/other
item?)
A *big* data shuffle? (ALL data would need to have its position
recalculated, with immediate end-of-the-world data shuffle?)
Nothing at all? (And the big fat warning is there only to
take fun on
the uninstructed ones? Not plausible...)
Give it a try! Download the CRUSHMap and run tests on it with
crushtool:
$ crushtool -i mycrushmap --test --rule 0 --num-rep 3
--show-statistics
Now, change the map, compile it and run again:
$ crushtool -i mycrushmap.new --test --rule 0 --num-rep 3
--show-statistics
Check the differences and you get the idea of how much has changed.
Wido
Thanks Wido ! :)
Thanks !
Loris
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
