Re: [Ceph-maintainers] Updated ceph backport
On Fri, Oct 17, 2014 at 1:02 AM, Gaudenz Steinlin <gaudenz@debian.org> wrote:
> Gregory Farnum <greg@inktank.com> writes:
>
>> On Thu, Oct 16, 2014 at 1:50 AM, Gaudenz Steinlin <gaudenz@debian.org> wrote:
>>> As google-perftools are optional for ceph, the best fix for this is
>>> probably to remove the dependency in the backport for the architectures
>>> not supported in wheezy. This simple patch would probably do it (not
>>> tested):
>>
>> FYI, this is probably not the most desirable solution. The main thing
>> we use gperftools for is the tcmalloc memory allocator; the default
>> malloc tends to have serious memory fragmentation problems under our
>> workload.
>> Not that it won't work, but virtual memory allocations measured in
>> tens of gigabytes are not unheard-of (versus application-allocated
>> memory of a few hundred megabytes). I imagine suboptimal malloc is
>> preferable to not having the Ceph packages, but having the perftools
>> packages is even better. :)
>
> I guess that the ceph package on these architectures is mostly used
> because it also builds the librbd package. It's mostly a question of
> having librbd available to be able to build qemu with rbd backend
> support. Is this also a concern in this case?
I don't have any specific knowledge about the clients, but I suspect
they're a lot friendlier than the server daemons are. Good point. :)
-Greg
Software Engineer #42 @ http://inktank.com | http://ceph.com
Reply to: