You're right that you could be fine above the warning threshold, you can also be in a bad state while being within the threshold. There is no one
size fits all for how much memory you need. The defaults and recommendations are there as a general guide to get you started. If you want to push the limits or stray from the recommendations, you really need to be testing to see what you can do (not in production,
but trying your best to duplicate the production environment and load).
How much memory per PG you need is going to be dependent on every aspect of your hardware, latency, load, etc. The only way to come up with that number is to actually test scenarios and find out for yourself. We're currently in the process of figuring out if we can reduce the amount of RAM in our hosts, but doing so slowly and deliberately testing everything we can think of along the way.
How much memory per PG you need is going to be dependent on every aspect of your hardware, latency, load, etc. The only way to come up with that number is to actually test scenarios and find out for yourself. We're currently in the process of figuring out if we can reduce the amount of RAM in our hosts, but doing so slowly and deliberately testing everything we can think of along the way.
 |
David Turner |
Cloud Operations Engineer |
StorageCraft
Technology Corporation 380 Data Drive Suite 300 | Draper | Utah | 84020 Office: 801.871.2760 | Mobile: 385.224.2943 |
| If you are not the intended recipient of this message or received it erroneously, please notify the sender and delete it, together with any attachments, and be advised that any dissemination or copying of this message is prohibited. |
________________________________________
From: Thomas HAMEL [hmlth@xxxxxxxxxx]
Sent: Tuesday, October 11, 2016 10:48 AM
To: David Turner; David Turner
Cc: Andrus, Brian Contractor; ceph-users@xxxxxxxxxxxxxx; Andrus, Brian Contractor; ceph-users@xxxxxxxxxxxxxx
Subject: Re: too many PGs per OSD (326 > max 300) warning when ALL PGs are 256
Le 11 octobre 2016 15:10:22 GMT+02:00, David Turner <david.turner@xxxxxxxxxxxxxxxx> a écrit :
>
>
>This is terrifying to think about increasing that threshold. You may
>have enough system resources while running healthy, but when
>recovering, especially during peering, your memory usage can more than
>double. If you have too many pgs per osd and/or too many osds per
>machine based on how much memory you have in each host you very well
>could end up in an eternal peering situation. Peering causes more
>memory usage, the Linux OOM killer is killing your osds causing more
>peering, and the only way to get out of the situation is running to the
>datacenter with buckets of RAM.
>
If the problem is memory, then I don't understand the point of the warning. If you have multiple osds sharing the same server, you can feel safe by being below the threshold while being at risk of not having enough ram if hundreds of pgs on multple osd starts to peer at the same time.
What am I missing here ?
What is the right amount of RAM for a pg, since that looks like the right metric (for this issue) ?
Thomas Hamel
>________________________________
>
>[cid:image7cfd24.JPG@49688796.47b58d1e]<https://storagecraft.com>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>
>Le 10 octobre 2016 23:00:33 GMT+02:00, David Turner
><david.turner@xxxxxxxxxxxxxxxx<mailto:david.turner@xxxxxxxxxxxxxxxx>> a
>écrit :
>
>The default it uses can be controlled in your ceph.conf file. The
>ceph-deploy tool is a generic ceph deployment tool which does not have
>presets for rados gateway deployments or other specific deployments.
>When creating pools you can specify the amount of pgs in them with the
>tool so that it doesn't use your defaults.
>
>You are correct that when creating a lot of pools that it is wiser to
>specify a smaller amount of pgs, but that is a manual step on your end
>and best to use the pgcalc tool to know what you should be aiming at.
>I haven't used ceph with rgw so I can't speak to optimal settings here,
>but I can only imagine that there are several tutorials and guides out
>there.
>
>________________________________
>
>[cid:image6e3427.JPG@497f9106.4eae2722]<https://storagecraft.com>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>
>________________________________
>From: Andrus, Brian Contractor
>[bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>]
>Sent: Monday, October 10, 2016 12:18 PM
>To: David Turner;
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: RE: too many PGs per OSD (326 > max 300) warning
>when ALL PGs are 256
>
>David,
>Thanks for the info. I am getting an understanding of how this works.
>Now I used the ceph-deploy tool to create the rgw pools. It seems then
>that the tool isn’t the best at creating the pools necessary for an rgw
>gateway as it made all of them the default sizes for pg_num/pgp_num
>Perhaps, then, it is wiser to have a very low default for those so the
>ceph-deploy tool doesn assign a large value to something that will
>merely hold control or other metadata?
>
>
>Brian Andrus
>ITACS/Research Computing
>Naval Postgraduate School
>Monterey, California
>voice: 831-656-6238
>
>
>
>From: David Turner [mailto:david.turner@xxxxxxxxxxxxxxxx]
>Sent: Monday, October 10, 2016 10:33 AM
>To: Andrus, Brian Contractor
><bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>>;
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: RE: too many PGs per OSD (326 > max 300) warning
>when ALL PGs are 256
>
>You have 11 pools with 256 pgs, 1 pool with 128 and 1 pool with 64...
>that's 3,008 pgs in your entire cluster. Multiply that number by your
>replica size and divide by how many OSDs you have in your cluster and
>you'll see what your average PGs per osd is. Based on the replica size
>you shared, that's a total number of 6,528 copies of PGs to be divided
>amongst the OSDS in your cluster. Your cluster will be in warning if
>that number is greater than 300 per OSD, like you're seeing. When
>designing your cluster and how many pools, pgs, and replica size you
>will be setting, please consult the pgcalc tool found here
>http://ceph.com/pgcalc/. You cannot reduce the number of PGs in a
>pool, so the easiest way to resolve this issue is mostly likely going
>to be destroying pools and recreating them with the proper number of
>PGs.
>
>The PG number should be based on what percentage of the data in your
>cluster will be in this pool. If I'm planning to have about 1024 PGs
>total in my cluster and I give 256 PGs to 4 different pools, then what
>I'm saying is that each of those 4 pools will have the exact same
>amount of data as each other. On the other hand, if I believe that 1
>of those pools will have 90% of the data and the other 3 pools will
>have very little data, then I'll probably give the larger pool 1024 PGs
>and the rest of them 64 PGs (or less depending on what I'm aiming for).
>It is beneficial to keep the pg_num and pgp_num counts to base 2
>numbers.
>________________________________
>[cid:image001.jpg@01D222E8.1655BB40]<https://storagecraft.com>
>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>________________________________
>From: ceph-users
>[ceph-users-bounces@xxxxxxxxxxxxxx<mailto:ceph-users-bounces@xxxxxxxxxxxxxx>]
>on behalf of Andrus, Brian Contractor
>[bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>]
>Sent: Monday, October 10, 2016 11:14 AM
>To: ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: too many PGs per OSD (326 > max 300) warning when
>ALL PGs are 256
>Ok, this is an odd one to me…
>I have several pools, ALL of them are set with pg_num and pgp_num =
>256. Yet, the warning about too many PGs per OSD is showing up.
>Here are my pools:
>
>pool 0 'rbd' replicated size 3 min_size 2 crush_ruleset 0 object_hash
>rjenkins pg_num 256 pgp_num 256 last_change 134 flags hashpspool
>stripe_width 0
>pool 1 'cephfs_data' replicated size 3 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 203 flags
>hashpspool crash_replay_interval 45 stripe_width 0
>pool 2 'cephfs_metadata' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 64 pgp_num 64 last_change 196 flags
>hashpspool stripe_width 0
>pool 3 'vmimages' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 128 pgp_num 128 last_change 213 flags
>hashpspool stripe_width 0
> removed_snaps [1~3]
>pool 25 '.rgw.root' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6199 flags
>hashpspool stripe_width 0
>pool 26 'default.rgw.control' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6202 flags hashpspool stripe_width 0
>pool 27 'default.rgw.data.root' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6204 flags hashpspool stripe_width 0
>pool 28 'default.rgw.gc' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6205 flags
>hashpspool stripe_width 0
>pool 29 'default.rgw.log' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6206 flags
>hashpspool stripe_width 0
>pool 30 'default.rgw.users.uid' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6211 flags hashpspool stripe_width 0
>pool 31 'default.rgw.meta' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6214 flags
>hashpspool stripe_width 0
>pool 32 'default.rgw.buckets.index' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6216 flags hashpspool stripe_width 0
>pool 33 'default.rgw.buckets.data' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6218 flags hashpspool stripe_width 0
>
>
>so why would the warning show up, and how do I get it to go away and
>stay away?
>
>
>Brian Andrus
>ITACS/Research Computing
>Naval Postgraduate School
>Monterey, California
>voice: 831-656-6238
>
>
>________________________________
>
>ceph-users mailing list
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>--
>Envoyé de mon appareil Android avec K-9 Mail. Veuillez excuser ma
>brièveté.
--
Envoyé de mon appareil Android avec K-9 Mail. Veuillez excuser ma brièveté.
From: Thomas HAMEL [hmlth@xxxxxxxxxx]
Sent: Tuesday, October 11, 2016 10:48 AM
To: David Turner; David Turner
Cc: Andrus, Brian Contractor; ceph-users@xxxxxxxxxxxxxx; Andrus, Brian Contractor; ceph-users@xxxxxxxxxxxxxx
Subject: Re: too many PGs per OSD (326 > max 300) warning when ALL PGs are 256
Le 11 octobre 2016 15:10:22 GMT+02:00, David Turner <david.turner@xxxxxxxxxxxxxxxx> a écrit :
>
>
>This is terrifying to think about increasing that threshold. You may
>have enough system resources while running healthy, but when
>recovering, especially during peering, your memory usage can more than
>double. If you have too many pgs per osd and/or too many osds per
>machine based on how much memory you have in each host you very well
>could end up in an eternal peering situation. Peering causes more
>memory usage, the Linux OOM killer is killing your osds causing more
>peering, and the only way to get out of the situation is running to the
>datacenter with buckets of RAM.
>
If the problem is memory, then I don't understand the point of the warning. If you have multiple osds sharing the same server, you can feel safe by being below the threshold while being at risk of not having enough ram if hundreds of pgs on multple osd starts to peer at the same time.
What am I missing here ?
What is the right amount of RAM for a pg, since that looks like the right metric (for this issue) ?
Thomas Hamel
>________________________________
>
>[cid:image7cfd24.JPG@49688796.47b58d1e]<https://storagecraft.com>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>
>Le 10 octobre 2016 23:00:33 GMT+02:00, David Turner
><david.turner@xxxxxxxxxxxxxxxx<mailto:david.turner@xxxxxxxxxxxxxxxx>> a
>écrit :
>
>The default it uses can be controlled in your ceph.conf file. The
>ceph-deploy tool is a generic ceph deployment tool which does not have
>presets for rados gateway deployments or other specific deployments.
>When creating pools you can specify the amount of pgs in them with the
>tool so that it doesn't use your defaults.
>
>You are correct that when creating a lot of pools that it is wiser to
>specify a smaller amount of pgs, but that is a manual step on your end
>and best to use the pgcalc tool to know what you should be aiming at.
>I haven't used ceph with rgw so I can't speak to optimal settings here,
>but I can only imagine that there are several tutorials and guides out
>there.
>
>________________________________
>
>[cid:image6e3427.JPG@497f9106.4eae2722]<https://storagecraft.com>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>
>________________________________
>From: Andrus, Brian Contractor
>[bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>]
>Sent: Monday, October 10, 2016 12:18 PM
>To: David Turner;
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: RE: too many PGs per OSD (326 > max 300) warning
>when ALL PGs are 256
>
>David,
>Thanks for the info. I am getting an understanding of how this works.
>Now I used the ceph-deploy tool to create the rgw pools. It seems then
>that the tool isn’t the best at creating the pools necessary for an rgw
>gateway as it made all of them the default sizes for pg_num/pgp_num
>Perhaps, then, it is wiser to have a very low default for those so the
>ceph-deploy tool doesn assign a large value to something that will
>merely hold control or other metadata?
>
>
>Brian Andrus
>ITACS/Research Computing
>Naval Postgraduate School
>Monterey, California
>voice: 831-656-6238
>
>
>
>From: David Turner [mailto:david.turner@xxxxxxxxxxxxxxxx]
>Sent: Monday, October 10, 2016 10:33 AM
>To: Andrus, Brian Contractor
><bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>>;
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: RE: too many PGs per OSD (326 > max 300) warning
>when ALL PGs are 256
>
>You have 11 pools with 256 pgs, 1 pool with 128 and 1 pool with 64...
>that's 3,008 pgs in your entire cluster. Multiply that number by your
>replica size and divide by how many OSDs you have in your cluster and
>you'll see what your average PGs per osd is. Based on the replica size
>you shared, that's a total number of 6,528 copies of PGs to be divided
>amongst the OSDS in your cluster. Your cluster will be in warning if
>that number is greater than 300 per OSD, like you're seeing. When
>designing your cluster and how many pools, pgs, and replica size you
>will be setting, please consult the pgcalc tool found here
>http://ceph.com/pgcalc/. You cannot reduce the number of PGs in a
>pool, so the easiest way to resolve this issue is mostly likely going
>to be destroying pools and recreating them with the proper number of
>PGs.
>
>The PG number should be based on what percentage of the data in your
>cluster will be in this pool. If I'm planning to have about 1024 PGs
>total in my cluster and I give 256 PGs to 4 different pools, then what
>I'm saying is that each of those 4 pools will have the exact same
>amount of data as each other. On the other hand, if I believe that 1
>of those pools will have 90% of the data and the other 3 pools will
>have very little data, then I'll probably give the larger pool 1024 PGs
>and the rest of them 64 PGs (or less depending on what I'm aiming for).
>It is beneficial to keep the pg_num and pgp_num counts to base 2
>numbers.
>________________________________
>[cid:image001.jpg@01D222E8.1655BB40]<https://storagecraft.com>
>
>David Turner | Cloud Operations Engineer | StorageCraft Technology
>Corporation<https://storagecraft.com>
>380 Data Drive Suite 300 | Draper | Utah | 84020
>Office: 801.871.2760 | Mobile: 385.224.2943
>
>________________________________
>If you are not the intended recipient of this message or received it
>erroneously, please notify the sender and delete it, together with any
>attachments, and be advised that any dissemination or copying of this
>message is prohibited.
>
>________________________________
>________________________________
>From: ceph-users
>[ceph-users-bounces@xxxxxxxxxxxxxx<mailto:ceph-users-bounces@xxxxxxxxxxxxxx>]
>on behalf of Andrus, Brian Contractor
>[bdandrus@xxxxxxx<mailto:bdandrus@xxxxxxx>]
>Sent: Monday, October 10, 2016 11:14 AM
>To: ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>Subject: too many PGs per OSD (326 > max 300) warning when
>ALL PGs are 256
>Ok, this is an odd one to me…
>I have several pools, ALL of them are set with pg_num and pgp_num =
>256. Yet, the warning about too many PGs per OSD is showing up.
>Here are my pools:
>
>pool 0 'rbd' replicated size 3 min_size 2 crush_ruleset 0 object_hash
>rjenkins pg_num 256 pgp_num 256 last_change 134 flags hashpspool
>stripe_width 0
>pool 1 'cephfs_data' replicated size 3 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 203 flags
>hashpspool crash_replay_interval 45 stripe_width 0
>pool 2 'cephfs_metadata' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 64 pgp_num 64 last_change 196 flags
>hashpspool stripe_width 0
>pool 3 'vmimages' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 128 pgp_num 128 last_change 213 flags
>hashpspool stripe_width 0
> removed_snaps [1~3]
>pool 25 '.rgw.root' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6199 flags
>hashpspool stripe_width 0
>pool 26 'default.rgw.control' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6202 flags hashpspool stripe_width 0
>pool 27 'default.rgw.data.root' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6204 flags hashpspool stripe_width 0
>pool 28 'default.rgw.gc' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6205 flags
>hashpspool stripe_width 0
>pool 29 'default.rgw.log' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6206 flags
>hashpspool stripe_width 0
>pool 30 'default.rgw.users.uid' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6211 flags hashpspool stripe_width 0
>pool 31 'default.rgw.meta' replicated size 2 min_size 1 crush_ruleset 0
>object_hash rjenkins pg_num 256 pgp_num 256 last_change 6214 flags
>hashpspool stripe_width 0
>pool 32 'default.rgw.buckets.index' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6216 flags hashpspool stripe_width 0
>pool 33 'default.rgw.buckets.data' replicated size 2 min_size 1
>crush_ruleset 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change
>6218 flags hashpspool stripe_width 0
>
>
>so why would the warning show up, and how do I get it to go away and
>stay away?
>
>
>Brian Andrus
>ITACS/Research Computing
>Naval Postgraduate School
>Monterey, California
>voice: 831-656-6238
>
>
>________________________________
>
>ceph-users mailing list
>ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
>http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>--
>Envoyé de mon appareil Android avec K-9 Mail. Veuillez excuser ma
>brièveté.
--
Envoyé de mon appareil Android avec K-9 Mail. Veuillez excuser ma brièveté.
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
