fuel-dev team mailing list archive

[Andrew Woodward <awoodward@xxxxxxxxxxxx>]

The journal logic is designed to divide the journal devices evenly among
the OSD devices on the system.

in 4.0 the logic was improved to create separate partitions inside the
journal device. The journal device will be a raw block device.

cases
1 journal, 1 osd = 1:1
1 journal, 2 osd = 2:1
2 journal, 2 osd = 1:1
2 journal, 5 osd = 2.5:1 (the first journal will end up with an extra osd)
4 journal, 20 osd = 5:1


On Fri, Jan 17, 2014 at 7:16 AM, Mike Scherbakov
<mscherbakov@xxxxxxxxxxxx>wrote:

> +fuel-dev
> On Jan 17, 2014 6:42 PM, "Roman Sokolkov" <rsokolkov@xxxxxxxxxxxx> wrote:
>
>> Andrey,
>>
>> are we able to use multiple block devices for journal purpose right now?
>>
>> Answer on your question is yes. I've checked it already. I've used 2
>> disks for journal and 2 for OSD. Single journal disk mapped on single OSD
>> disk in this case. I assume if i will add additional (third) OSD disk, it
>> will remain without separate journal. Here is the logic<https://github.com/stackforge/fuel-library/blob/master/deployment/puppet/ceph/lib/facter/ceph_osd.rb#L30>
>> .
>>
>> But the main question was:
>>
>> Is there ability to use one journal disk for multiple OSD disks?
>>
>> i.e. i have 4 SSDs and 20 HDDs in storage node. I want to use 1 SSD per 5
>> HDD. In this case SSD should contain 5 partitions.
>>
>> Thanks, Roman S.
>>
>>
>> On Fri, Jan 17, 2014 at 6:31 PM, Andrey Korolev <akorolev@xxxxxxxxxxxx>wrote:
>>
>>> Hello,
>>>
>>> Services team came with a question about Ceph setup - are we able to use
>>> multiple block devices for journal purpose right now?
>>>
>>> Also I am reviving back the issue with journal placement - our current
>>> approach to slice journal device to the equal chunk is not very good in
>>> terms of possible reconfiguration - e.g. you can not add more OSD daemons
>>> without shutting down every daemon on the selected nodes, doing journal
>>> flush and repartitioning. Can we please switch to the filesystem-based
>>> journal placement in the next release? Passing 'discard' option to the
>>> filesystem we will be able to maintain same level of wearout and with fixed
>>> or proportional journal size (e.g. to not allocate all free space) we will
>>> achieve greater maintenance flexibility.
>>>
>>
>>
>>
>> --
>> Best regards, Roman Sokolkov
>>
>