An algorithm to fix uneven CRUSH distributions in Ceph

The current CRUSH implementation in Ceph does not always provide an even distribution.

The most common cause of unevenness is when only a few thousands PGs, or less, are mapped. This is not enough samples and the variations can be as high as 25%. For instance, when there are two OSDs with the same weight, distributing randomly four PGs among them may lead to one OSD having three PGs and the other only one. This problem would be resolved by having at least thousands of PGs per OSD, but that is not recommended because it would require too many resources.

The other cause of uneven distribution is conditional probability. For a two-replica pool, PGs are mapped to OSDs that must be different: the second OSD is chosen at random, on the condition that it is not the same as the first OSD. When all OSDs have the same probability, this bias is not significant. But when OSDs have different weights it makes a difference. For instance, given nine OSDs with weight 1 and one OSD with weight 5, the smaller OSDs will be overfilled (from 7% to 10%) and the bigger OSD will be ~15% underfilled.

The proposed algorithm fixes both cases by producing new weights that can be used as a weight set in Luminous clusters.

For a given pool the input parameters are:

  • pool size
  • numeric id
  • number of PGs
  • root of the CRUSH rule (take step)
  • the CRUSH rule itself
   for size in [1,pool size]
     copy all weights to the size - 1 weight set
     recursively walk the root
     for each bucket at a given level in the hierarchy
       repeat until the difference with the expected distribution is small
         map all PGs in the pool, with size instead of pool size
         in the size - 1 weight set
           lower the weight of the most overfilled child
           increase  the weight of the most underfilled child

It is common to change the size of a pool in a Ceph cluster. When increasing the size from 2 to 3, the user expects the existing objects to stay where they are, with new objects being created to provide an additional replica. To preserve this property while optimizing the weights, there needs to be a different weight set for each possible size. This is what the outer loop (for size in [1,pool size]) does.

If the distribution is not as expected at the highest level of the hierarchy, there is no way to fix that at the lowest levels. For instance if a host receives 100 more PGs that it should, the OSDs it contains will inevitably be overfilled. This is why the optimization proceeds from the top of the hierarchy.

When a bucket is given precisely the expected number of PGs and fails to distribute them evenly among its children, the children’s weights can be modified to get closer to the ideal distribution. Increasing the weight of the most underfilled item will capture PGs from the other buckets. And decreasing the weight of the most overfilled will push PGs out of it. A simulation is run to determine precisely which PGs will be distributed to which item because there is no known mathematical formula to calculate that. This is why all PGs are mapped to determine which items are over- or underfilled.

Continue reading “An algorithm to fix uneven CRUSH distributions in Ceph”

Ceph full ratio and uneven CRUSH distributions

A common CRUSH rule in Ceph is

    step chooseleaf firstn 0 type host

meaning Placement Groups (PGs) will place replicas on different hosts so the cluster can sustain the failure of any host without losing data. The missing replicas are then restored from the surviving replicas (via a process called “backfilling”) and placed on the remaining hosts.

To make sure there is enough space in the cluster to cope with the failure of a host, a certain percentage of free space in the cluster is reserved (from the beginning) and never used. This percentage needs to be adjusted to take into account the most overfull OSD in case the PG distribution is not even.

For instance, in a cluster with five hosts containing two identical disks each, reserving 20% plus 6.45% to account for the most overfull OSD displayed below should be enough:

crush analyze --type device --rule data \
              --replication-count 2 \
              --crushmap mymap.txt \
              --pool 0 --pg-num 1024 --pgp-num 1024
         ~id~  ~weight~  ~objects~  ~over/under used %~
~name~
device9     9       1.0        218                 6.45
device5     5       1.0        214                 4.49
device7     7       1.0        214                 4.49
device0     0       1.0        212                 3.52
device8     8       1.0        212                 3.52
device6     6       1.0        208                 1.56
device2     2       1.0        201                -1.86
device3     3       1.0        201                -1.86
device4     4       1.0        192                -6.25
device1     1       1.0        176               -14.06

However this uneven distribution will be different when a host is removed, because that causes a change in the most overfull OSD – and the new most overfull OSD may even be worse (more overfull) than the previous one. In our example cluster, device9 was the most (6.45%) overfull. If the host containing device8 and device9 is removed, device5 becomes the most overfull OSD, and it is worse (8.98%):

         ~id~  ~weight~  ~objects~  ~over/under used %~
~name~
device5     5       1.0        279                 8.98
device7     7       1.0        270                 5.47
device2     2       1.0        268                 4.69
device0     0       1.0        267                 4.30
device3     3       1.0        249                -2.73
device6     6       1.0        246                -3.91
device1     1       1.0        241                -5.86
device4     4       1.0        228               -10.94

It would therefore be better to reserve 28.98% instead of 26.45% to make sure the cluster does not become too full after a host failure. To help with that, the crush analyze command was modified to display the worst case scenario for each bucket type in the crushmap:

crush analyze --type device --rule data \
              --replication-count 2 \
              --crushmap mymap.txt \
              --pool 0 --pg-num 1024 --pgp-num 1024
         ~id~  ~weight~  ~objects~  ~over/under used %~
~name~
device9     9       1.0        218                 6.45
device5     5       1.0        214                 4.49
device7     7       1.0        214                 4.49
device0     0       1.0        212                 3.52
device8     8       1.0        212                 3.52
device6     6       1.0        208                 1.56
device2     2       1.0        201                -1.86
device3     3       1.0        201                -1.86
device4     4       1.0        192                -6.25
device1     1       1.0        176               -14.06

Worst case scenario if a host fails:

        ~over used %~
~type~
device           8.98
host             4.49
root             0.00

Improving PGs distribution with CRUSH weight sets

In a Ceph cluster with a single pool of 1024 Placement Groups (PGs), the PG distribution among devices will not be as expected. (see Predicting Ceph PG placement for details about this uneven distribution).

In the following, the difference between the expected number of PGs on a given host or device and the actual number of PGs can be as high as 25%. For instance, device4 has a weight of 1 and is expected to get 128 PGs but it only gets 96. And host2 (containing device4) has 83 less PGs than expected.

         ~expected~  ~actual~  ~delta~   ~delta%~  ~weight~
dc1            1024      1024        0   0.000000       1.0
 host0          256       294       38  14.843750       2.0
  device0       128       153       25  19.531250       1.0
  device1       128       141       13  10.156250       1.0
 host1          256       301       45  17.578125       2.0
  device2       128       157       29  22.656250       1.0
  device3       128       144       16  12.500000       1.0
 host2          512       429      -83 -16.210938       4.0
  device4       128        96      -32 -25.000000       1.0
  device5       128       117      -11  -8.593750       1.0
  device6       256       216      -40 -15.625000       2.0

Using minimization methods it is possible to find alternate weights for hosts and devices that significantly reduce the uneven distribution, as shown below.

         ~expected~  ~actual~  ~delta~  ~delta%~  ~weight~
dc1            1024      1024        0  0.000000  1.000000
 host0          256       259        3  1.171875  0.680399
  device0       128       129        1  0.781250  0.950001
  device1       128       130        2  1.562500  1.050000
 host1          256       258        2  0.781250  0.646225
  device2       128       129        1  0.781250  0.867188
  device3       128       129        1  0.781250  1.134375
 host2          512       507       -5 -0.976562  7.879046
  device4       128       126       -2 -1.562500  1.111111
  device5       128       127       -1 -0.781250  0.933333
  device6       256       254       -2 -0.781250  1.955555

For each host or device, we now have two weights. First, there is the target weight which is used to calculate the expected number of PGs. For instance, host0 should get 1024 * (target weight host0 = 2 / (target weight host0 = 2 + target weight host1 = 2 + target weight host3 = 4)) PGs, that is 1024 * ( 2 / ( 2 + 2 + 4 ) ) = 256. The second weight is weight set, which is used to choose where a PGs is mapped.

The problem is that the crushmap format only allows one weight per host or device. Although it would be possible to store the target weight outside of the crushmap, it would be inconvenient and difficult to understand. Instead, the crushmap syntax was extended to allow additional weights (via weight set) to be associated with each item (i.e. disk or host in the example below).

                                                    target
         ~expected~  ~actual~  ~delta~   ~delta%~  ~weight~ ~weight set~
dc1            1024      1024        0   0.000000       1.0     1.000000
 host0          256       294       38  14.843750       2.0     0.680399
  device0       128       153       25  19.531250       1.0     0.950001
  device1       128       141       13  10.156250       1.0     1.050000
 host1          256       301       45  17.578125       2.0     0.646225
  device2       128       157       29  22.656250       1.0     0.867188
  device3       128       144       16  12.500000       1.0     1.134375
 host2          512       429      -83 -16.210938       4.0     7.879046
  device4       128        96      -32 -25.000000       1.0     1.111111
  device5       128       117      -11  -8.593750       1.0     0.933333
  device6       256       216      -40 -15.625000       2.0     1.955555

Continue reading “Improving PGs distribution with CRUSH weight sets”

Faster Ceph CRUSH computation with smaller buckets

The CRUSH function maps Ceph placement groups (PGs) and objects to OSDs. It is used extensively in Ceph clients and daemons as well as in the Linux kernel modules and its CPU cost should be reduced to the minimum.

It is common to define the Ceph CRUSH map so that PGs use OSDs on different hosts. The hierarchy can be as simple as 100 hosts, each containing 7 OSDs. When determining which OSDs belong to a given PG, in this scenario the CRUSH function will hash each of the 100 hosts to the PG, compare the results, and select the OSDs that score the highest. That is 100 + 7 calls to the hash function.

If the hierarchy is modified to have 20 racks, each containing 5 hosts with 7 OSDs per host, we only have 20 + 5 + 7 calls to the hash function. The first 20 calls are to select the rack, the next 5 to select the host in the rack and the last 7 to select the OSD within the host.

In other words, this hierarchical subdivision of the CRUSH map reduces the number of calls to the hash function from 107 to 32. It does not have any influence on how the PGs are mapped to OSDs but it saves CPU.