Maybe changing the number of concurrent back fills could limit the memory
usage.
On 9 Jan 2016 05:52, "Josef Johansson" <[email protected]> wrote:

> Hi,
>
> I would say this is normal. 1GB of ram per 1TB is what we designed the
> cluster for, I would believe that an EC-pool demands a lot more. Buy more
> ram and start everything 32GB ram is quite little, when the cluster is
> operating OK you'll see that extra ram getting used as file cache which
> makes the cluster faster.
>
> Regards,
> Josef
> On 6 Jan 2016 12:12, "Kenneth Waegeman" <[email protected]> wrote:
>
>> Hi all,
>>
>> We experienced some serious trouble with our cluster: A running cluster
>> started failing and started a chain reaction until the ceph cluster was
>> down, as about half the OSDs are down (in a EC pool)
>>
>> Each host has 8 OSDS of 8 TB (i.e. RAID 0 of 2 4TB disk) for an EC pool
>> (10+3, 14 hosts) and 2 cache OSDS and 32 GB of RAM.
>> The reason we have the Raid0 of the disks, is because we tried with 16
>> disk before, but 32GB didn't seem enough to keep the cluster stable
>>
>> We don't know for sure what triggered the chain reaction, but what we
>> certainly see, is that while recovering, our OSDS are using a lot of
>> memory. We've seen some OSDS using almost 8GB of RAM (resident; virtual
>> 11GB)
>> So right now we don't have enough memory to recover the cluster, because
>> the  OSDS  get killed by OOMkiller before they can recover..
>> And I don't know doubling our memory will be enough..
>>
>> A few questions:
>>
>> * Does someone has seen this before?
>> * 2GB was still normal, but 8GB seems a lot, is this expected behaviour?
>> * We didn't see this with an nearly empty cluster. Now it was filled
>> about 1/4 (270TB). I guess it would become worse when filled half or more?
>> * How high can this memory usage become ? Can we calculate the maximum
>> memory of an OSD? Can we limit it ?
>> * We can upgrade/reinstall to infernalis, will that solve anything?
>>
>> This is related to a previous post of me :
>> http://permalink.gmane.org/gmane.comp.file-systems.ceph.user/22259
>>
>>
>> Thank you very much !!
>>
>> Kenneth
>>
>> _______________________________________________
>> ceph-users mailing list
>> [email protected]
>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>>
>
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com

Reply via email to