After some more work i realized that didn't get me closer at all. It was still
only selecting 13 osds *and* still occasionally re-selecting the same one. I
think the multiple emit/takes isn't working like I expect. Given:
step take default
step chooseleaf indep 0 type host
step emit
step take default
step chooseleaf indep 0 type host
step emit
In a rule, I would expect it to try to select ~1 osd per host once, and then
start over again. Instead, what I'm seeing is it selects ~1 osd per host and
then when it starts again, it re-selects those same osds, resulting in multiple
placements on 2 or 3 osds per pg.
It turns out what I'm trying to do is described here:
https://www.mail-archive.com/ceph-users%40lists.ceph.com/msg01076.html
But I can't find any other references to anything like this.
Thanks, Aaron
> On Dec 23, 2014, at 9:23 AM, Aaron Bassett wrote:
>
> I’m trying to set up an erasure coded pool with k=9 m=6 on 13 osd hosts. I’m
> trying to write a crush rule for this which will balance this between hosts
> as much as possible. I understand that having 9+6=15 > 13, I will need to
> parse the tree twice in order to find enough pgs. So what I’m trying to do is
> select ~1 from each host on the first pass, and then select n more osds to
> fill it out, without using any osds from the first pass, and preferably
> balancing them between racks.
>
> For starters, I don't know if this is even possible or if its the right
> approach to what I'm trying to do, but heres my attempt:
>
> rule .us-phx.rgw.buckets.ec {
>ruleset 1
>type erasure
>min_size 3
>max_size 20
>step set_chooseleaf_tries 5
>step take default
>step chooseleaf indep 0 type host
>step emit
>step take default
>step chooseleaf indep 0 type rack
>step emit
> }
>
> This gets me pretty close, the first pass works great and the second pass
> does a nice balance between racks, but in my testing ~ 6 out of 1000 pgs will
> have two osds in their group. I'm guessing I need to get down to one pass to
> make sure that doesn't happen, but I'm having a hard time sorting out how to
> hit the requirement of balancing among hosts *and* allowing for more than one
> osd per host.
>
> Thanks, Aaron
___
ceph-users mailing list
ceph-users@lists.ceph.com
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com