7 Autoscaling placement groups
8 ============================
10 Placement groups (PGs) are an internal implementation detail of how
11 Ceph distributes data. You can allow the cluster to either make
12 recommendations or automatically tune PGs based on how the cluster is
13 used by enabling *pg-autoscaling*.
15 Each pool in the system has a ``pg_autoscale_mode`` property that can be set to ``off``, ``on``, or ``warn``.
17 * ``off``: Disable autoscaling for this pool. It is up to the administrator to choose an appropriate PG number for each pool. Please refer to :ref:`choosing-number-of-placement-groups` for more information.
18 * ``on``: Enable automated adjustments of the PG count for the given pool.
19 * ``warn``: Raise health alerts when the PG count should be adjusted
21 To set the autoscaling mode for existing pools,::
23 ceph osd pool set <pool-name> pg_autoscale_mode <mode>
25 For example to enable autoscaling on pool ``foo``,::
27 ceph osd pool set foo pg_autoscale_mode on
29 You can also configure the default ``pg_autoscale_mode`` that is
30 applied to any pools that are created in the future with::
32 ceph config set global osd_pool_default_pg_autoscale_mode <mode>
34 Viewing PG scaling recommendations
35 ----------------------------------
37 You can view each pool, its relative utilization, and any suggested changes to
38 the PG count with this command::
40 ceph osd pool autoscale-status
42 Output will be something like::
44 POOL SIZE TARGET SIZE RATE RAW CAPACITY RATIO TARGET RATIO EFFECTIVE RATIO PG_NUM NEW PG_NUM AUTOSCALE
45 a 12900M 3.0 82431M 0.4695 8 128 warn
46 c 0 3.0 82431M 0.0000 0.2000 0.9884 1 64 warn
47 b 0 953.6M 3.0 82431M 0.0347 8 warn
49 **SIZE** is the amount of data stored in the pool. **TARGET SIZE**, if
50 present, is the amount of data the administrator has specified that
51 they expect to eventually be stored in this pool. The system uses
52 the larger of the two values for its calculation.
54 **RATE** is the multiplier for the pool that determines how much raw
55 storage capacity is consumed. For example, a 3 replica pool will
56 have a ratio of 3.0, while a k=4,m=2 erasure coded pool will have a
59 **RAW CAPACITY** is the total amount of raw storage capacity on the
60 OSDs that are responsible for storing this pool's (and perhaps other
61 pools') data. **RATIO** is the ratio of that total capacity that
62 this pool is consuming (i.e., ratio = size * rate / raw capacity).
64 **TARGET RATIO**, if present, is the ratio of storage that the
65 administrator has specified that they expect this pool to consume
66 relative to other pools with target ratios set.
67 If both target size bytes and ratio are specified, the
68 ratio takes precedence.
70 **EFFECTIVE RATIO** is the target ratio after adjusting in two ways:
72 1. subtracting any capacity expected to be used by pools with target size set
73 2. normalizing the target ratios among pools with target ratio set so
74 they collectively target the rest of the space. For example, 4
75 pools with target_ratio 1.0 would have an effective ratio of 0.25.
77 The system uses the larger of the actual ratio and the effective ratio
80 **PG_NUM** is the current number of PGs for the pool (or the current
81 number of PGs that the pool is working towards, if a ``pg_num``
82 change is in progress). **NEW PG_NUM**, if present, is what the
83 system believes the pool's ``pg_num`` should be changed to. It is
84 always a power of 2, and will only be present if the "ideal" value
85 varies from the current value by more than a factor of 3.
87 The final column, **AUTOSCALE**, is the pool ``pg_autoscale_mode``,
88 and will be either ``on``, ``off``, or ``warn``.
94 Allowing the cluster to automatically scale PGs based on usage is the
95 simplest approach. Ceph will look at the total available storage and
96 target number of PGs for the whole system, look at how much data is
97 stored in each pool, and try to apportion the PGs accordingly. The
98 system is relatively conservative with its approach, only making
99 changes to a pool when the current number of PGs (``pg_num``) is more
100 than 3 times off from what it thinks it should be.
102 The target number of PGs per OSD is based on the
103 ``mon_target_pg_per_osd`` configurable (default: 100), which can be
106 ceph config set global mon_target_pg_per_osd 100
108 The autoscaler analyzes pools and adjusts on a per-subtree basis.
109 Because each pool may map to a different CRUSH rule, and each rule may
110 distribute data across different devices, Ceph will consider
111 utilization of each subtree of the hierarchy independently. For
112 example, a pool that maps to OSDs of class `ssd` and a pool that maps
113 to OSDs of class `hdd` will each have optimal PG counts that depend on
114 the number of those respective device types.
117 .. _specifying_pool_target_size:
119 Specifying expected pool size
120 -----------------------------
122 When a cluster or pool is first created, it will consume a small
123 fraction of the total cluster capacity and will appear to the system
124 as if it should only need a small number of placement groups.
125 However, in most cases cluster administrators have a good idea which
126 pools are expected to consume most of the system capacity over time.
127 By providing this information to Ceph, a more appropriate number of
128 PGs can be used from the beginning, preventing subsequent changes in
129 ``pg_num`` and the overhead associated with moving data around when
130 those adjustments are made.
132 The *target size* of a pool can be specified in two ways: either in
133 terms of the absolute size of the pool (i.e., bytes), or as a weight
134 relative to other pools with a ``target_size_ratio`` set.
138 ceph osd pool set mypool target_size_bytes 100T
140 will tell the system that `mypool` is expected to consume 100 TiB of
141 space. Alternatively,::
143 ceph osd pool set mypool target_size_ratio 1.0
145 will tell the system that `mypool` is expected to consume 1.0 relative
146 to the other pools with ``target_size_ratio`` set. If `mypool` is the
147 only pool in the cluster, this means an expected use of 100% of the
148 total capacity. If there is a second pool with ``target_size_ratio``
149 1.0, both pools would expect to use 50% of the cluster capacity.
151 You can also set the target size of a pool at creation time with the optional ``--target-size-bytes <bytes>`` or ``--target-size-ratio <ratio>`` arguments to the ``ceph osd pool create`` command.
153 Note that if impossible target size values are specified (for example,
154 a capacity larger than the total cluster) then a health warning
155 (``POOL_TARGET_SIZE_BYTES_OVERCOMMITTED``) will be raised.
157 If both ``target_size_ratio`` and ``target_size_bytes`` are specified
158 for a pool, only the ratio will be considered, and a health warning
159 (``POOL_HAS_TARGET_SIZE_BYTES_AND_RATIO``) will be issued.
161 Specifying bounds on a pool's PGs
162 ---------------------------------
164 It is also possible to specify a minimum number of PGs for a pool.
165 This is useful for establishing a lower bound on the amount of
166 parallelism client will see when doing IO, even when a pool is mostly
167 empty. Setting the lower bound prevents Ceph from reducing (or
168 recommending you reduce) the PG number below the configured number.
170 You can set the minimum number of PGs for a pool with::
172 ceph osd pool set <pool-name> pg_num_min <num>
174 You can also specify the minimum PG count at pool creation time with
175 the optional ``--pg-num-min <num>`` argument to the ``ceph osd pool
180 A preselection of pg_num
181 ========================
183 When creating a new pool with::
185 ceph osd pool create {pool-name} [pg_num]
187 it is optional to choose the value of ``pg_num``. If you do not
188 specify ``pg_num``, the cluster can (by default) automatically tune it
189 for you based on how much data is stored in the pool (see above, :ref:`pg-autoscaler`).
191 Alternatively, ``pg_num`` can be explicitly provided. However,
192 whether you specify a ``pg_num`` value or not does not affect whether
193 the value is automatically tuned by the cluster after the fact. To
194 enable or disable auto-tuning,::
196 ceph osd pool set {pool-name} pg_autoscale_mode (on|off|warn)
198 The "rule of thumb" for PGs per OSD has traditionally be 100. With
199 the additional of the balancer (which is also enabled by default), a
200 value of more like 50 PGs per OSD is probably reasonable. The
201 challenge (which the autoscaler normally does for you), is to:
203 - have the PGs per pool proportional to the data in the pool, and
204 - end up with 50-100 PGs per OSDs, after the replication or
205 erasuring-coding fan-out of each PG across OSDs is taken into
208 How are Placement Groups used ?
209 ===============================
211 A placement group (PG) aggregates objects within a pool because
212 tracking object placement and object metadata on a per-object basis is
213 computationally expensive--i.e., a system with millions of objects
214 cannot realistically track placement on a per-object basis.
217 /-----\ /-----\ /-----\ /-----\ /-----\
218 | obj | | obj | | obj | | obj | | obj |
219 \-----/ \-----/ \-----/ \-----/ \-----/
221 +--------+--------+ +---+----+
224 +-----------------------+ +-----------------------+
225 | Placement Group #1 | | Placement Group #2 |
227 +-----------------------+ +-----------------------+
229 +------------------------------+
232 +-----------------------+
235 +-----------------------+
237 The Ceph client will calculate which placement group an object should
238 be in. It does this by hashing the object ID and applying an operation
239 based on the number of PGs in the defined pool and the ID of the pool.
240 See `Mapping PGs to OSDs`_ for details.
242 The object's contents within a placement group are stored in a set of
243 OSDs. For instance, in a replicated pool of size two, each placement
244 group will store objects on two OSDs, as shown below.
247 +-----------------------+ +-----------------------+
248 | Placement Group #1 | | Placement Group #2 |
250 +-----------------------+ +-----------------------+
253 /----------\ /----------\ /----------\ /----------\
255 | OSD #1 | | OSD #2 | | OSD #2 | | OSD #3 |
257 \----------/ \----------/ \----------/ \----------/
260 Should OSD #2 fail, another will be assigned to Placement Group #1 and
261 will be filled with copies of all objects in OSD #1. If the pool size
262 is changed from two to three, an additional OSD will be assigned to
263 the placement group and will receive copies of all objects in the
266 Placement groups do not own the OSD; they share it with other
267 placement groups from the same pool or even other pools. If OSD #2
268 fails, the Placement Group #2 will also have to restore copies of
269 objects, using OSD #3.
271 When the number of placement groups increases, the new placement
272 groups will be assigned OSDs. The result of the CRUSH function will
273 also change and some objects from the former placement groups will be
274 copied over to the new Placement Groups and removed from the old ones.
276 Placement Groups Tradeoffs
277 ==========================
279 Data durability and even distribution among all OSDs call for more
280 placement groups but their number should be reduced to the minimum to
288 After an OSD fails, the risk of data loss increases until the data it
289 contained is fully recovered. Let's imagine a scenario that causes
290 permanent data loss in a single placement group:
292 - The OSD fails and all copies of the object it contains are lost.
293 For all objects within the placement group the number of replica
294 suddenly drops from three to two.
296 - Ceph starts recovery for this placement group by choosing a new OSD
297 to re-create the third copy of all objects.
299 - Another OSD, within the same placement group, fails before the new
300 OSD is fully populated with the third copy. Some objects will then
301 only have one surviving copies.
303 - Ceph picks yet another OSD and keeps copying objects to restore the
304 desired number of copies.
306 - A third OSD, within the same placement group, fails before recovery
307 is complete. If this OSD contained the only remaining copy of an
308 object, it is permanently lost.
310 In a cluster containing 10 OSDs with 512 placement groups in a three
311 replica pool, CRUSH will give each placement groups three OSDs. In the
312 end, each OSDs will end up hosting (512 * 3) / 10 = ~150 Placement
313 Groups. When the first OSD fails, the above scenario will therefore
314 start recovery for all 150 placement groups at the same time.
316 The 150 placement groups being recovered are likely to be
317 homogeneously spread over the 9 remaining OSDs. Each remaining OSD is
318 therefore likely to send copies of objects to all others and also
319 receive some new objects to be stored because they became part of a
322 The amount of time it takes for this recovery to complete entirely
323 depends on the architecture of the Ceph cluster. Let say each OSD is
324 hosted by a 1TB SSD on a single machine and all of them are connected
325 to a 10Gb/s switch and the recovery for a single OSD completes within
326 M minutes. If there are two OSDs per machine using spinners with no
327 SSD journal and a 1Gb/s switch, it will at least be an order of
330 In a cluster of this size, the number of placement groups has almost
331 no influence on data durability. It could be 128 or 8192 and the
332 recovery would not be slower or faster.
334 However, growing the same Ceph cluster to 20 OSDs instead of 10 OSDs
335 is likely to speed up recovery and therefore improve data durability
336 significantly. Each OSD now participates in only ~75 placement groups
337 instead of ~150 when there were only 10 OSDs and it will still require
338 all 19 remaining OSDs to perform the same amount of object copies in
339 order to recover. But where 10 OSDs had to copy approximately 100GB
340 each, they now have to copy 50GB each instead. If the network was the
341 bottleneck, recovery will happen twice as fast. In other words,
342 recovery goes faster when the number of OSDs increases.
344 If this cluster grows to 40 OSDs, each of them will only host ~35
345 placement groups. If an OSD dies, recovery will keep going faster
346 unless it is blocked by another bottleneck. However, if this cluster
347 grows to 200 OSDs, each of them will only host ~7 placement groups. If
348 an OSD dies, recovery will happen between at most of ~21 (7 * 3) OSDs
349 in these placement groups: recovery will take longer than when there
350 were 40 OSDs, meaning the number of placement groups should be
353 No matter how short the recovery time is, there is a chance for a
354 second OSD to fail while it is in progress. In the 10 OSDs cluster
355 described above, if any of them fail, then ~17 placement groups
356 (i.e. ~150 / 9 placement groups being recovered) will only have one
357 surviving copy. And if any of the 8 remaining OSD fail, the last
358 objects of two placement groups are likely to be lost (i.e. ~17 / 8
359 placement groups with only one remaining copy being recovered).
361 When the size of the cluster grows to 20 OSDs, the number of Placement
362 Groups damaged by the loss of three OSDs drops. The second OSD lost
363 will degrade ~4 (i.e. ~75 / 19 placement groups being recovered)
364 instead of ~17 and the third OSD lost will only lose data if it is one
365 of the four OSDs containing the surviving copy. In other words, if the
366 probability of losing one OSD is 0.0001% during the recovery time
367 frame, it goes from 17 * 10 * 0.0001% in the cluster with 10 OSDs to 4 * 20 *
368 0.0001% in the cluster with 20 OSDs.
370 In a nutshell, more OSDs mean faster recovery and a lower risk of
371 cascading failures leading to the permanent loss of a Placement
372 Group. Having 512 or 4096 Placement Groups is roughly equivalent in a
373 cluster with less than 50 OSDs as far as data durability is concerned.
375 Note: It may take a long time for a new OSD added to the cluster to be
376 populated with placement groups that were assigned to it. However
377 there is no degradation of any object and it has no impact on the
378 durability of the data contained in the Cluster.
380 .. _object distribution:
382 Object distribution within a pool
383 ---------------------------------
385 Ideally objects are evenly distributed in each placement group. Since
386 CRUSH computes the placement group for each object, but does not
387 actually know how much data is stored in each OSD within this
388 placement group, the ratio between the number of placement groups and
389 the number of OSDs may influence the distribution of the data
392 For instance, if there was a single placement group for ten OSDs in a
393 three replica pool, only three OSD would be used because CRUSH would
394 have no other choice. When more placement groups are available,
395 objects are more likely to be evenly spread among them. CRUSH also
396 makes every effort to evenly spread OSDs among all existing Placement
399 As long as there are one or two orders of magnitude more Placement
400 Groups than OSDs, the distribution should be even. For instance, 256
401 placement groups for 3 OSDs, 512 or 1024 placement groups for 10 OSDs
404 Uneven data distribution can be caused by factors other than the ratio
405 between OSDs and placement groups. Since CRUSH does not take into
406 account the size of the objects, a few very large objects may create
407 an imbalance. Let say one million 4K objects totaling 4GB are evenly
408 spread among 1024 placement groups on 10 OSDs. They will use 4GB / 10
409 = 400MB on each OSD. If one 400MB object is added to the pool, the
410 three OSDs supporting the placement group in which the object has been
411 placed will be filled with 400MB + 400MB = 800MB while the seven
412 others will remain occupied with only 400MB.
416 Memory, CPU and network usage
417 -----------------------------
419 For each placement group, OSDs and MONs need memory, network and CPU
420 at all times and even more during recovery. Sharing this overhead by
421 clustering objects within a placement group is one of the main reasons
424 Minimizing the number of placement groups saves significant amounts of
427 .. _choosing-number-of-placement-groups:
429 Choosing the number of Placement Groups
430 =======================================
432 .. note: It is rarely necessary to do this math by hand. Instead, use the ``ceph osd pool autoscale-status`` command in combination with the ``target_size_bytes`` or ``target_size_ratio`` pool properties. See :ref:`pg-autoscaler` for more information.
434 If you have more than 50 OSDs, we recommend approximately 50-100
435 placement groups per OSD to balance out resource usage, data
436 durability and distribution. If you have less than 50 OSDs, choosing
437 among the `preselection`_ above is best. For a single pool of objects,
438 you can use the following formula to get a baseline
440 Total PGs = :math:`\frac{OSDs \times 100}{pool \: size}`
442 Where **pool size** is either the number of replicas for replicated
443 pools or the K+M sum for erasure coded pools (as returned by **ceph
444 osd erasure-code-profile get**).
446 You should then check if the result makes sense with the way you
447 designed your Ceph cluster to maximize `data durability`_,
448 `object distribution`_ and minimize `resource usage`_.
450 The result should always be **rounded up to the nearest power of two**.
452 Only a power of two will evenly balance the number of objects among
453 placement groups. Other values will result in an uneven distribution of
454 data across your OSDs. Their use should be limited to incrementally
455 stepping from one power of two to another.
457 As an example, for a cluster with 200 OSDs and a pool size of 3
458 replicas, you would estimate your number of PGs as follows
460 :math:`\frac{200 \times 100}{3} = 6667`. Nearest power of 2: 8192
462 When using multiple data pools for storing objects, you need to ensure
463 that you balance the number of placement groups per pool with the
464 number of placement groups per OSD so that you arrive at a reasonable
465 total number of placement groups that provides reasonably low variance
466 per OSD without taxing system resources or making the peering process
469 For instance a cluster of 10 pools each with 512 placement groups on
470 ten OSDs is a total of 5,120 placement groups spread over ten OSDs,
471 that is 512 placement groups per OSD. That does not use too many
472 resources. However, if 1,000 pools were created with 512 placement
473 groups each, the OSDs will handle ~50,000 placement groups each and it
474 would require significantly more resources and time for peering.
476 You may find the `PGCalc`_ tool helpful.
479 .. _setting the number of placement groups:
481 Set the Number of Placement Groups
482 ==================================
484 To set the number of placement groups in a pool, you must specify the
485 number of placement groups at the time you create the pool.
486 See `Create a Pool`_ for details. Even after a pool is created you can also change the number of placement groups with::
488 ceph osd pool set {pool-name} pg_num {pg_num}
490 After you increase the number of placement groups, you must also
491 increase the number of placement groups for placement (``pgp_num``)
492 before your cluster will rebalance. The ``pgp_num`` will be the number of
493 placement groups that will be considered for placement by the CRUSH
494 algorithm. Increasing ``pg_num`` splits the placement groups but data
495 will not be migrated to the newer placement groups until placement
496 groups for placement, ie. ``pgp_num`` is increased. The ``pgp_num``
497 should be equal to the ``pg_num``. To increase the number of
498 placement groups for placement, execute the following::
500 ceph osd pool set {pool-name} pgp_num {pgp_num}
502 When decreasing the number of PGs, ``pgp_num`` is adjusted
503 automatically for you.
505 Get the Number of Placement Groups
506 ==================================
508 To get the number of placement groups in a pool, execute the following::
510 ceph osd pool get {pool-name} pg_num
513 Get a Cluster's PG Statistics
514 =============================
516 To get the statistics for the placement groups in your cluster, execute the following::
518 ceph pg dump [--format {format}]
520 Valid formats are ``plain`` (default) and ``json``.
523 Get Statistics for Stuck PGs
524 ============================
526 To get the statistics for all placement groups stuck in a specified state,
527 execute the following::
529 ceph pg dump_stuck inactive|unclean|stale|undersized|degraded [--format <format>] [-t|--threshold <seconds>]
531 **Inactive** Placement groups cannot process reads or writes because they are waiting for an OSD
532 with the most up-to-date data to come up and in.
534 **Unclean** Placement groups contain objects that are not replicated the desired number
535 of times. They should be recovering.
537 **Stale** Placement groups are in an unknown state - the OSDs that host them have not
538 reported to the monitor cluster in a while (configured by ``mon_osd_report_timeout``).
540 Valid formats are ``plain`` (default) and ``json``. The threshold defines the minimum number
541 of seconds the placement group is stuck before including it in the returned statistics
542 (default 300 seconds).
548 To get the placement group map for a particular placement group, execute the following::
556 Ceph will return the placement group map, the placement group, and the OSD status::
558 osdmap e13 pg 1.6c (1.6c) -> up [1,0] acting [1,0]
564 To retrieve statistics for a particular placement group, execute the following::
566 ceph pg {pg-id} query
569 Scrub a Placement Group
570 =======================
572 To scrub a placement group, execute the following::
574 ceph pg scrub {pg-id}
576 Ceph checks the primary and any replica nodes, generates a catalog of all objects
577 in the placement group and compares them to ensure that no objects are missing
578 or mismatched, and their contents are consistent. Assuming the replicas all
579 match, a final semantic sweep ensures that all of the snapshot-related object
580 metadata is consistent. Errors are reported via logs.
582 To scrub all placement groups from a specific pool, execute the following::
584 ceph osd pool scrub {pool-name}
586 Prioritize backfill/recovery of a Placement Group(s)
587 ====================================================
589 You may run into a situation where a bunch of placement groups will require
590 recovery and/or backfill, and some particular groups hold data more important
591 than others (for example, those PGs may hold data for images used by running
592 machines and other PGs may be used by inactive machines/less relevant data).
593 In that case, you may want to prioritize recovery of those groups so
594 performance and/or availability of data stored on those groups is restored
595 earlier. To do this (mark particular placement group(s) as prioritized during
596 backfill or recovery), execute the following::
598 ceph pg force-recovery {pg-id} [{pg-id #2}] [{pg-id #3} ...]
599 ceph pg force-backfill {pg-id} [{pg-id #2}] [{pg-id #3} ...]
601 This will cause Ceph to perform recovery or backfill on specified placement
602 groups first, before other placement groups. This does not interrupt currently
603 ongoing backfills or recovery, but causes specified PGs to be processed
604 as soon as possible. If you change your mind or prioritize wrong groups,
607 ceph pg cancel-force-recovery {pg-id} [{pg-id #2}] [{pg-id #3} ...]
608 ceph pg cancel-force-backfill {pg-id} [{pg-id #2}] [{pg-id #3} ...]
610 This will remove "force" flag from those PGs and they will be processed
611 in default order. Again, this doesn't affect currently processed placement
612 group, only those that are still queued.
614 The "force" flag is cleared automatically after recovery or backfill of group
617 Similarly, you may use the following commands to force Ceph to perform recovery
618 or backfill on all placement groups from a specified pool first::
620 ceph osd pool force-recovery {pool-name}
621 ceph osd pool force-backfill {pool-name}
625 ceph osd pool cancel-force-recovery {pool-name}
626 ceph osd pool cancel-force-backfill {pool-name}
628 to restore to the default recovery or backfill priority if you change your mind.
630 Note that these commands could possibly break the ordering of Ceph's internal
631 priority computations, so use them with caution!
632 Especially, if you have multiple pools that are currently sharing the same
633 underlying OSDs, and some particular pools hold data more important than others,
634 we recommend you use the following command to re-arrange all pools's
635 recovery/backfill priority in a better order::
637 ceph osd pool set {pool-name} recovery_priority {value}
639 For example, if you have 10 pools you could make the most important one priority 10,
640 next 9, etc. Or you could leave most pools alone and have say 3 important pools
641 all priority 1 or priorities 3, 2, 1 respectively.
646 If the cluster has lost one or more objects, and you have decided to
647 abandon the search for the lost data, you must mark the unfound objects
650 If all possible locations have been queried and objects are still
651 lost, you may have to give up on the lost objects. This is
652 possible given unusual combinations of failures that allow the cluster
653 to learn about writes that were performed before the writes themselves
656 Currently the only supported option is "revert", which will either roll back to
657 a previous version of the object or (if it was a new object) forget about it
658 entirely. To mark the "unfound" objects as "lost", execute the following::
660 ceph pg {pg-id} mark_unfound_lost revert|delete
662 .. important:: Use this feature with caution, because it may confuse
663 applications that expect the object(s) to exist.
673 .. _Create a Pool: ../pools#createpool
674 .. _Mapping PGs to OSDs: ../../../architecture#mapping-pgs-to-osds
675 .. _pgcalc: http://ceph.com/pgcalc/