right pg_num value for CephFS Quick Start guide
Hello, While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result. About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that disk size might have a role to play, I changed the number of OSDs to 2 each with 20 GB disks and later with 50 GB disks but neither helped. I used dnf to install ceph and ceph-deploy to setup the cluster. I've copied the the cluster status after every attempt here[2] in case that helps. Any suggestions about pg_num values I should choose and on the pg_num values that would be nice for a user looking forward to get quickly started with CephFS? [1] https://docs.ceph.com/docs/master/start/quick-cephfs/ [2] https://paste.fedoraproject.org/paste/Q-WH8VWtwu6JwF7eW2JmnA Thanks, - Rishabh
On Wed, Sep 11, 2019 at 10:38 AM Rishabh Dave <ridave@redhat.com> wrote:
Hello,
While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result.
About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that disk size might have a role to play, I changed the number of OSDs to 2 each with 20 GB disks and later with 50 GB disks but neither helped. I used dnf to install ceph and ceph-deploy to setup the cluster.
I've copied the the cluster status after every attempt here[2] in case that helps. Any suggestions about pg_num values I should choose and on the pg_num values that would be nice for a user looking forward to get quickly started with CephFS?
Why not recommend this in quick-start for master or from nautilus stable? https://ceph.com/rados/new-in-nautilus-pg-merging-and-autotuning/
[1] https://docs.ceph.com/docs/master/start/quick-cephfs/ [2] https://paste.fedoraproject.org/paste/Q-WH8VWtwu6JwF7eW2JmnA
Thanks, - Rishabh _______________________________________________ Dev mailing list -- dev@ceph.io To unsubscribe send an email to dev-leave@ceph.io
On Wed, 11 Sep 2019 at 23:13, Vasu Kulkarni <vakulkar@redhat.com> wrote:
On Wed, Sep 11, 2019 at 10:38 AM Rishabh Dave <ridave@redhat.com> wrote:
Hello,
While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result.
About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that disk size might have a role to play, I changed the number of OSDs to 2 each with 20 GB disks and later with 50 GB disks but neither helped. I used dnf to install ceph and ceph-deploy to setup the cluster.
I've copied the the cluster status after every attempt here[2] in case that helps. Any suggestions about pg_num values I should choose and on the pg_num values that would be nice for a user looking forward to get quickly started with CephFS?
Why not recommend this in quick-start for master or from nautilus stable? https://ceph.com/rados/new-in-nautilus-pg-merging-and-autotuning/
Although, I could get "active+clean" for all PGs but not HEALTH_OK, it'll mention be great to mention this anyway. Thanks for pointing out.
On Wed, 11 Sep 2019, Rishabh Dave wrote:
Hello,
While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result.
About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that
This is unrelated to the PGs or the capacity--the problem is that you have a single node, and the default CRUSH rule replicates across hosts. That's why your pools are unhealthy. You can fix this by creating a new crush rule with 'osd' instead of 'host' as the failure domain, and then setting your pool(s) to use that rule. osd crush rule create-replicated <name> <root> create crush rule <name> for replicated pool to <type> {<class>} start from <root>, replicate across buckets of type <type>, use devices of type <class> (ssd or hdd) osd pool set <poolname> crush_rule <rule-name> sage
disk size might have a role to play, I changed the number of OSDs to 2 each with 20 GB disks and later with 50 GB disks but neither helped. I used dnf to install ceph and ceph-deploy to setup the cluster.
I've copied the the cluster status after every attempt here[2] in case that helps. Any suggestions about pg_num values I should choose and on the pg_num values that would be nice for a user looking forward to get quickly started with CephFS?
[1] https://docs.ceph.com/docs/master/start/quick-cephfs/ [2] https://paste.fedoraproject.org/paste/Q-WH8VWtwu6JwF7eW2JmnA
Thanks, - Rishabh _______________________________________________ Dev mailing list -- dev@ceph.io To unsubscribe send an email to dev-leave@ceph.io
On Wed, 11 Sep 2019 at 23:31, Sage Weil <sage@newdream.net> wrote:
On Wed, 11 Sep 2019, Rishabh Dave wrote:
Hello,
While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result.
About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that
This is unrelated to the PGs or the capacity--the problem is that you have a single node, and the default CRUSH rule replicates across hosts. That's why your pools are unhealthy.
You can fix this by creating a new crush rule with 'osd' instead of 'host' as the failure domain, and then setting your pool(s) to use that rule.
osd crush rule create-replicated <name> <root> create crush rule <name> for replicated pool to <type> {<class>} start from <root>, replicate across buckets of type <type>, use devices of type <class> (ssd or hdd)
osd pool set <poolname> crush_rule <rule-name>
sage
Both, setting the new crush rule to pools and creating OSDs on separate nodes worked. Thanks! However, in both cases the PG status wasn't "active+clean". It was "54 active+undersized" and "10 active+undersized+degraded". IMO, it would ideal for CephFS Quick Start guide to lead to "active+clean". Is there anything more that can be done and would be suitable for the guide?
Hi, if you have the proper setup you should always reach active+clean for all your PGs, - Single node with 2 OSDs: Rule replicates across OSD, set size=2 and min_size=1 on your pool - Single node with 3 OSDs: Rule replicates across OSD (default will be size=3 min_size=2 on your pool) - Multiple nodes with 2 nodes: Rule replicates across HOST, set size=2 and min_size=1 on your pool - Multiple nodes with3 nodes: Rule replicates across HOST (default will be size=3 min_size=2 on your pool) Tip: For single node deployment set osd_crush_chooseleaf_type = 0 in your configuration file [global] section before you deploy your MONs and OSDs and it will create the correct CRUSH rule Regards JC
On Sep 13, 2019, at 09:57, Rishabh Dave <ridave@redhat.com> wrote:
On Wed, 11 Sep 2019 at 23:31, Sage Weil <sage@newdream.net <mailto:sage@newdream.net>> wrote:
On Wed, 11 Sep 2019, Rishabh Dave wrote:
Hello,
While working on CephFS Quick Start guide[1], the major issue that I came across was choosing the value for pg_num for the pools that will serve CephFS. I've tried the values from 4 to 128 for both data and metadata pools and have always got "undersized+peered" instead of "active+clean". Copying pg_num values from the cluster setup by vstart.sh (8 for data and 16 for metadata pools) gave me the same result.
About the cluster: I had a single node running Fedora 29 with 1 MON, 1 MGR, 1 MDS and 3 OSDs each with a disk size of 10 GB. Thinking that
This is unrelated to the PGs or the capacity--the problem is that you have a single node, and the default CRUSH rule replicates across hosts. That's why your pools are unhealthy.
You can fix this by creating a new crush rule with 'osd' instead of 'host' as the failure domain, and then setting your pool(s) to use that rule.
osd crush rule create-replicated <name> <root> create crush rule <name> for replicated pool to <type> {<class>} start from <root>, replicate across buckets of type <type>, use devices of type <class> (ssd or hdd)
osd pool set <poolname> crush_rule <rule-name>
sage
Both, setting the new crush rule to pools and creating OSDs on separate nodes worked. Thanks!
However, in both cases the PG status wasn't "active+clean". It was "54 active+undersized" and "10 active+undersized+degraded". IMO, it would ideal for CephFS Quick Start guide to lead to "active+clean". Is there anything more that can be done and would be suitable for the guide? _______________________________________________ Dev mailing list -- dev@ceph.io <mailto:dev@ceph.io> To unsubscribe send an email to dev-leave@ceph.io <mailto:dev-leave@ceph.io>
On Sat, 14 Sep 2019 at 00:46, JC Lopez <jelopez@redhat.com> wrote:
Hi,
if you have the proper setup you should always reach active+clean for all your PGs,
- Single node with 2 OSDs: Rule replicates across OSD, set size=2 and min_size=1 on your pool - Single node with 3 OSDs: Rule replicates across OSD (default will be size=3 min_size=2 on your pool) - Multiple nodes with 2 nodes: Rule replicates across HOST, set size=2 and min_size=1 on your pool - Multiple nodes with3 nodes: Rule replicates across HOST (default will be size=3 min_size=2 on your pool)
Tip: For single node deployment set osd_crush_chooseleaf_type = 0 in your configuration file [global] section before you deploy your MONs and OSDs and it will create the correct CRUSH rule
I get a perfect cluster status now. I can see HEALTH_OK as well "active+clean" for all PGs. Thank you!
participants (4)
-
JC Lopez
-
Rishabh Dave
-
Sage Weil
-
Vasu Kulkarni