"Advanced power loss protection" is in fact a performance feature, not a safety
one.
28 февраля 2019 г. 13:03:51 GMT+03:00, Uwe Sauter <uwe.sauter...@gmail.com>
пишет:
>Hi all,
>
>thanks for your insights.
>
>Eneko,
>
>> We tried to use a Samsung 840 Pro SSD as OSD some time ago and it was
>a no-go; it wasn't that performance was bad, it
>> just didn't work for the kind of use of OSD. Any HDD was better than
>it (the disk was healthy and have been used in a
>> software raid-1 for a pair of years).
>>
>> I suggest you check first that your Samsung 860 Pro disks work well
>for Ceph. Also, how is your host's RAM?
>
>As already mentioned the hosts each have 64GB RAM. Each host has 3 SSDs
>for OSD usage. Each OSD is using about 1.3GB virtual
>memory / 400MB residual memory.
>
>
>
>Joachim,
>
>> I can only recommend the use of enterprise SSDs. We've tested many
>consumer SSDs in the past, including your SSDs. Many
>> of them are not suitable for long-term use and some weard out within
>6 months.
>
>Unfortunately I couldn't afford enterprise grade SSDs. But I suspect
>that my workload (about 20 VMs for our infrastructure, the
>most IO demanding is probably LDAP) is light enough that wearout won't
>be a problem.
>
>The issue I'm seeing then is probably related to direct IO if using
>bluestore. But with filestore, the file system cache probably
>hides the latency issues.
>
>
>Igor,
>
>> AFAIR Samsung 860 Pro isn't for enterprise market, you shouldn't use
>consumer SSDs for Ceph.
>>
>> I had some experience with Samsung 960 Pro a while ago and it turned
>out that it handled fsync-ed writes very slowly
>> (comparing to the original/advertised performance). Which one can
>probably explain by the lack of power loss protection
>> for these drives. I suppose it's the same in your case.
>>
>> Here are a couple links on the topic:
>>
>>
>https://www.percona.com/blog/2018/02/08/fsync-performance-storage-devices/
>>
>>
>https://www.sebastien-han.fr/blog/2014/10/10/ceph-how-to-test-if-your-ssd-is-suitable-as-a-journal-device/
>
>Power loss protection wasn't a criteria for me as the cluster hosts are
>distributed in two buildings with separate battery backed
>UPSs. As mentioned above I suspect the main difference for my case
>between filestore and bluestore is file system cache vs. direct
>IO. Which means I will keep using filestore.
>
>Regards,
>
> Uwe
>_______________________________________________
>ceph-users mailing list
>ceph-users@lists.ceph.com
>http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
--
With best regards,
Vitaliy Filippov
_______________________________________________
ceph-users mailing list
ceph-users@lists.ceph.com
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
