Skip to content

Pool Operations

A DAOS pool is a storage reservation that can span any number of storage engines in a DAOS system. Pools are managed by the administrator. The amount of space allocated to a pool is decided at creation time with the dmg pool create command. Pools can be expanded at a later time with the dmg pool expand command that adds additional engine ranks to the existing pool's storage allocation. The DAOS management API also provides these capabilities.

Pool Basics

The dmg pool command is the main administrative tool to manage pools. Its subcommands can be grouped into the following areas:

  • Commands to create, list, query, extend, and destroy a pool. These are the basic functions to manage the storage allocation in a pool.

  • Commands to list and set pool properties, and commands to set and list Access Control Lists (ACLs) for a pool.

  • Commands to manage failures and other non-standard scenarios. This includes draining, excluding and re-integrating targets, and evicting client connections to a pool.

  • An upgrade command to upgrade a pool's format version after a DAOS software update to a newer DAOS version.

Creating a Pool

A DAOS pool can be created through the dmg pool create command. The mandatory parameters that are needed for the creation of a pool are the pool label, and a specification of the size of the storage allocation.

The pool label must consist of alphanumeric characters, colon (:), period (.), hyphen (-) or underscore (_). The maximum length of a pool label is 127 characters. Labels that can be parsed as a UUID (e.g. 123e4567-e89b-12d3-a456-426614174000) are forbidden. Pool labels must be unique across the DAOS system.

A pool's size is determined by two factors: How many storage engines are participating in the storage allocation, and how much capacity in each storage tier is allocated (the latter can be specified either on a per-engine basis, or as the total for the pool across all participating engines). The same amount of storage will be allocated on each of the participating storage engines. If one or more of those engines do not have sufficient free space for the requested capacity, the pool creation will fail.

If neither the --nranks nor the --ranks option is used, then the pool will span all storage engines of the DAOS system. To limit the pool to only a subset of the engines, those two options can be used to specify either the desired number of engines, or an explicit list of engine ranks to be used for this pool.

The capacity of the pool can be specified in three different ways:

  1. The --size option can be used to specify the total pool capacity in Bytes. This value denotes the sum of the SCM and NVMe capacities. The relative contributions of the SCM and NVMe storage tiers to the total pool size are determined by the --tier-ratio parameter. By default this ratio is 6,94, so for a pool of size 100TB there will be 6TB of SCM and 94 TB of NVMe storage. An SCM-only pool can be created by using --tier-ratio 100,0.

  2. The --size option can be used to specify the total pool capacity as a percentage of the currently free capacity. In this case, the tier ratio will be ignored. For example, requesting --size=100% will allocate 100% of the free SCM capacity and 100% of the free NVMe capacity to the pool, regardless of the ratio of those two free capacity values.

  3. This implies that it is not possible to create an SCM-only pool by using a percentage size (unless there is no NVMe storage in the system at all, and all pools are SCM-only).

  4. If the amount of free space is different across the participating engines, then the minimum free space is used to calculate the space that is allocated per engine.

  5. Because the percentage numbers refer to currently free space and not total space, the absolute size of a pool created with --size=percentage% will be impacted by other concurrent pool create operations. The command output will always list the total capacities in addition to the requested percentage.

  6. The --scm-size parameter (and optionally --nvme-size) can be used to specify the SCM capacity (and optionally the NVMe capacity) per storage engine in Bytes. The minimum SCM size is 16 MiB per target, so for a storage engine with 16 targets the minimum is --scm-size=256MiB. The NVMe size can be zero. If it is non-zero then the minimum NVMe size is 1 GiB per target, so for a storage engine with 16 targets the minimum non-zero NVMe size is --nvme-size=16GiB. To derive the total pool capacity, these per-engine capacities have to be multiplied by the number of participating engines.

Note

The suffixes "M", "MB", "G", "GB", "T" or "TB" denote base-10 capacities, whereas "MiB", "GiB" or "TiB" denote base-2. So in the first example above, specifying --scm-size=256GB would fail as 256 GB is smaller than the minimum 256 GiB.

Examples:

To create a pool labeled tank:

$ dmg pool create --size=${N}TB tank

This command creates a pool labeled tank distributed across the DAOS servers with a target size on each server that is comprised of $N * 0.94 TB of NVMe storage and $N * 0.06 TB of SCM storage. The default SCM:NVMe ratio may be adjusted at pool creation time as described above.

The UUID allocated to the newly created pool is printed to stdout as well as the pool service replica ranks.

$ dmg pool create --help
...

[create command options]
      -g, --group=      DAOS pool to be owned by given group, format name@domain
      -u, --user=       DAOS pool to be owned by given user, format name@domain
      -p, --label=      Unique label for pool (deprecated, use positional argument)
      -P, --properties= Pool properties to be set
      -a, --acl-file=   Access Control List file path for DAOS pool
      -z, --size=       Total size of DAOS pool (auto)
      -t, --tier-ratio= Percentage of storage tiers for pool storage (auto) (default: 6% SCM, 94% NVMe)
      -k, --nranks=     Number of ranks to use (auto)
      -v, --nsvc=       Number of pool service replicas
      -s, --scm-size=   Per-engine SCM allocation for DAOS pool (manual)
      -n, --nvme-size=  Per-engine NVMe allocation for DAOS pool (manual)
      -r, --ranks=      Storage engine unique identifiers (ranks) for DAOS pool

The typical output of this command is as follows:

$ dmg pool create --size 50GB tank
Creating DAOS pool with automatic storage allocation: 50 GB total, 6,94 tier ratio
Pool created with 6.00% SCM/NVMe ratio
-----------------------------------------
  UUID                 : 8a05bf3a-a088-4a77-bb9f-df989fce7cc8
  Service Ranks        : [1-3]
  Storage Ranks        : [0-15]
  Total Size           : 50 GB
  Storage tier 0 (SCM) : 3.0 GB (188 MB / rank)
  Storage tier 1 (NVMe): 47 GB (3.0 GB / rank)

This created a pool with UUID 8a05bf3a-a088-4a77-bb9f-df989fce7cc8, with pool service redundancy enabled by default (pool service replicas on ranks 1-3).

If no redundancy is desired, use --nsvc=1 in order to specify that only a single pool service replica should be created.

Note that it is difficult to determine the usable space by the user, and currently we cannot provide the precise value. The usable space depends not only on pool size, but also on number of targets, target size, object class, storage redundancy factor, etc.

Listing Pools

To see a list of the pools in the DAOS system:

$ dmg pool list
Pool     Size   Used Imbalance Disabled
----     ----   ---- --------- --------
tank     47 GB  0%   0%        0/32

This returns a table of pool labels (or UUIDs if no label was specified) with the following information for each pool: - the total pool size - the percentage of used space (i.e., 100 * used space / total space) - the imbalance percentage indicating whether data distribution across the difference storage targets is well balanced. 0% means that there is no imbalance and 100% means that out-of-space errors might be returned by some storage targets while space is still available on others. - the number of disabled targets (0 here) and the number of targets that the pool was originally configured with (total).

The --verbose option provides more detailed information including the number of service replicas, the full UUIDs and space distribution between SCM and NVMe for each pool:

$ dmg pool list --verbose
Label UUID                                 SvcReps SCM Size SCM Used SCM Imbalance NVME Size NVME Used NVME Imbalance Disabled
----- ----                                 ------- -------- -------- ------------- --------- --------- -------------- --------
tank  8a05bf3a-a088-4a77-bb9f-df989fce7cc8 1-3      3 GB    10 kB    0%            47 GB     0 B       0%             0/32

Destroying a Pool

To destroy a pool labeled tank:

$ dmg pool destroy tank
Pool-destroy command succeeded

The pool's UUID can be used instead of the pool label.

Querying a Pool

The pool query operation retrieves information (i.e., the number of targets, space usage, rebuild status, property list, and more) about an existing pool.

To query a pool labeled tank:

$ dmg pool query tank

The pool's UUID can be used instead of the pool label. Below is the output for a pool created with SCM space only.

    pool=47293abe-aa6f-4147-97f6-42a9f796d64a
    Pool 47293abe-aa6f-4147-97f6-42a9f796d64a, ntarget=64, disabled=8
    Pool space info:
    - Target(VOS) count:56
    - SCM:
        Total size: 28GB
        Free: 28GB, min:505MB, max:512MB, mean:512MB
    - NVMe:
        Total size: 0
        Free: 0, min:0, max:0, mean:0
    Rebuild done, 10 objs, 1026 recs

The total and free sizes are the sum across all the targets whereas min/max/mean gives information about individual targets. A min value close to 0 means that one target is running out of space.

NB: the Versioning Object Store (VOS) may reserve a portion of the SCM and NVMe allocations to mitigate against fragmentation and for background operations (e.g., aggregation, garbage collection). The amount of storage set aside depends on the size of the target and may take up 2+ GB. Therefore, out of space conditions may occur even while pool query may not show the minimum free space approaching zero.

The example below shows a rebuild in progress and NVMe space allocated.

    pool=95886b8b-7eb8-454d-845c-fc0ae0ba5671
    Pool 95886b8b-7eb8-454d-845c-fc0ae0ba5671, ntarget=64, disabled=8
    Pool space info:
    - Target(VOS) count:56
    - SCM:
        Total size: 28GB
        Free: 28GB, min:470MB, max:512MB, mean:509MB
    - NVMe:
        Total size: 56GB
        Free: 28GB, min:470MB, max:512MB, mean:509MB
    Rebuild busy, 75 objs, 9722 recs

Additional status and telemetry data is planned to be exported through management tools and will be documented here once available.

Evicting Users

To evict handles/connections to a pool labeled tank:

$ dmg pool evict tank
Pool-evict command succeeded

The pool's UUID can be used instead of the pool label.

Pool Properties

Properties are predefined parameters that the administrator can tune to control the behavior of a pool.

Properties Management

Current properties of an existing pool can be retrieved via the dmg pool get-prop command line.

$ dmg pool get-prop tank
Pool 8a05bf3a-a088-4a77-bb9f-df989fce7cc8 properties:
Name                            Value
----                            -----
EC cell size (ec_cell_sz)       1.0 MiB
Pool label (label)              tank
Reclaim strategy (reclaim)      lazy
Self-healing policy (self_heal) exclude
Rebuild space ratio (space_rb)  0%

All properties can be specified when creating the pool.

$ dmg pool create --size 50GB --properties reclaim:disabled tank2
Creating DAOS pool with automatic storage allocation: 50 GB NVMe + 6.00% SCM
Pool created with 100.00% SCM/NVMe ratio
-----------------------------------------
  UUID          : 1f265216-5877-4302-ad29-aa0f90df3f86
  Service Ranks : 0
  Storage Ranks : [0-1]
  Total Size    : 50 GB
  SCM           : 50 GB (25 GB / rank)
  NVMe          : 0 B (0 B / rank)

$ dmg pool get-prop tank2
Pool 1f265216-5877-4302-ad29-aa0f90df3f86 properties:
Name                            Value
----                            -----
EC cell size (ec_cell_sz)       1.0 MiB
Pool label (label)              tank2
Reclaim strategy (reclaim)      disabled
Self-healing policy (self_heal) exclude
Rebuild space ratio (space_rb)  0%

Some properties can be modified after pool creation via the set-prop option.

$ dmg pool set-prop tank2 reclaim:lazy
pool set-prop succeeded

$ dmg pool get-prop tank2 reclaim
Pool 1f265216-5877-4302-ad29-aa0f90df3f86 properties:
Name                       Value
----                       -----
Reclaim strategy (reclaim) lazy

Reclaim Strategy (reclaim)

DAOS is a versioned object store that tags every I/O with an epoch number. This versioning mechanism is the baseline for multi-version concurrency control and snapshot support. Over time, unused versions need to be reclaimed in order to release storage space and also simplify the metadata index. This process is called aggregation.

The reclaim property defines what strategy to use to reclaimed unused version. Three options are supported:

  • "lazy" : Trigger aggregation only when there is no IO activities or SCM free space is under pressure (default strategy)
  • "time" : Trigger aggregation regularly despite of IO activities.
  • "disabled" : Never trigger aggregation. The system will eventually run out of space even if data is being deleted.

Self-healing Policy (self_heal)

This property defines whether a failing engine is automatically evicted from the pool. Once excluded, the self-healing mechanism will be triggered to restore the pool data redundancy on the surviving storage engines. Two options are supported: "exclude" (default strategy) and "rebuild".

Reserved Space for Rebuilds (space_rb)

This property defines the percentage of total space reserved on each storage engine for self-healing purpose. The reserved space cannot be consumed by applications. Valid values are 0% to 100%, the default is 0%. When setting this property, specifying the percentage symbol is optional: space_rb:2% and space_rb:2 both specify two percent of storage capacity.

Default EC Cell Size (ec_cell_sz)

This property defines the default erasure code cell size inherited to DAOS containers. The EC cell size can be between 1kiB and 1GiB, although it should typically be set to a value between 32kiB and 1MiB. The default is 1MiB. When setting this property, the cell size can be specified in Bytes (as a number with no suffix), with a base-10 suffix like k or MB, or with a base-2 suffix like ki or MiB.

Access Control Lists

Client user and group access for pools are controlled by Access Control Lists (ACLs). Most pool-related tasks are performed using the DMG administrative tool, which is authenticated by the administrative certificate rather than user-specific credentials.

Access-controlled client pool accesses include:

  • Connecting to the pool.

  • Querying the pool.

  • Creating containers in the pool.

  • Deleting containers in the pool.

This is reflected in the set of supported pool permissions.

A user must be able to connect to the pool in order to access any containers inside, regardless of their permissions on those containers.

Ownership

By default, the dmg pool create command will use the current user and current group to set the pool's owner-user and owner-group. This default can be changed with the --user and --group options.

Pool ownership conveys no special privileges for access control decisions. All desired privileges of the owner-user (OWNER@) and owner-group (GROUP@) must be explicitly defined by an administrator in the pool ACL.

ACL at Pool Creation

To create a pool with a custom ACL:

$ dmg pool create --size <size> --acl-file <path> <pool_label>

The ACL file format is detailed in here.

Displaying ACL

To view a pool's ACL:

$ dmg pool get-acl --outfile=<path> <pool_label>

The output is in the same string format used in the ACL file during creation, with one Access Control Entry (i.e., ACE) per line.

An example output is presented below:

$ dmg pool get-acl tank
# Owner: jlombard@
# Owner Group: jlombard@
# Entries:
A::OWNER@:rw
A::bob@:r
A:G:GROUP@:rw

Modifying ACL

For all of these commands using an ACL file, the ACL file must be in the format noted above for container creation.

Overwriting ACL

To replace a pool's ACL with a new ACL:

$ dmg pool overwrite-acl --acl-file <path> <pool_label>

Adding and Updating ACEs

To add or update multiple entries in an existing pool ACL:

$ dmg pool update-acl --acl-file <path> <pool_label>

To add or update a single entry in an existing pool ACL:

$ dmg pool update-acl --entry <ACE> <pool_label>

If there is no existing entry for the principal in the ACL, the new entry is added to the ACL. If there is already an entry for the principal, that entry is replaced with the new one.

For instance:

$ dmg pool update-acl -e A::kelsey@:r tank
Pool-update-ACL command succeeded, UUID: 8a05bf3a-a088-4a77-bb9f-df989fce7cc8
# Owner: jlombard@
# Owner Group: jlombard@
# Entries:
A::OWNER@:rw
A::bob@:r
A::kelsey@:r
A:G:GROUP@:rw

Removing an ACE

To delete an entry for a given principal in an existing pool ACL:

$ dmg pool delete-acl --principal <principal> <pool_label>

The principal corresponds to the principal portion of an ACE that was set during pool creation or a previous pool ACL operation. For the delete operation, the principal argument must be formatted as follows:

  • Named user: u:username@
  • Named group: g:groupname@
  • Special principals: OWNER@, GROUP@, and EVERYONE@

The entry for that principal will be completely removed. This does not always mean that the principal will have no access. Rather, their access to the pool will be decided based on the remaining ACL rules.

Pool Modifications

Automatic Exclusion

An engine detected as dead by the SWIM monitoring protocol will, by default, be automatically excluded from all the pools using this engine. The engine will thus not only be marked as excluded by the system (i.e., in dmg system query), but also reported as disabled in the pool query output (i.e., dmg pool query) for all the impacted pools.

Upon exclusion, the collective rebuild process (i.e., also called self-healing) will be automatically triggered to restore data redundancy on the surviving engine.

Note

The rebuild process may consume many resources on each engine and is thus throttled to reduce the impact on application performance. This current logic relies on CPU cycles on the storage nodes. By default, the rebuild process is configured to consume up to 30% of the CPU cycles, leaving the other 70% for regular I/O operations.

Manual Exclusion

An operator can exclude one or more engines or targets from a specific DAOS pool using the rank the target resides, as well as the target idx on that rank. If a target idx list is not provided, all targets on the engine rank will be excluded.

To exclude a target from a pool:

$ dmg pool exclude --rank=${rank} --target-idx=${idx1},${idx2},${idx3} <pool_label>

The pool target exclude command accepts 2 parameters:

  • The engine rank of the target(s) to be excluded.
  • The target indices of the targets to be excluded from that engine rank (optional).

Upon successful manual exclusion, the self-healing mechanism will be triggered to restore redundancy on the remaining engines.

Drain

Alternatively, when an operator would like to remove one or more engines or targets without the system operating in degraded mode, the drain operation can be used. A pool drain operation initiates rebuild without excluding the designated engine or target until after the rebuild is complete. This allows the drained entity to continue to perform I/O while the rebuild operation is ongoing. Drain additionally enables non-replicated data to be rebuilt onto another target whereas in a conventional failure scenario non-replicated data would not be integrated into a rebuild and would be lost.

To drain a target from a pool:

$ dmg pool drain --rank=${rank} --target-idx=${idx1},${idx2},${idx3} $DAOS_POOL

The pool target drain command accepts 2 parameters:

  • The engine rank of the target(s) to be drained.
  • The target indices of the targets to be drained from that engine rank (optional).

Reintegration

After an engine failure and exclusion, an operator can fix the underlying issue and reintegrate the affected engines or targets to restore the pool to its original state. The operator can either reintegrate specific targets for an engine rank by supplying a target idx list, or reintegrate an entire engine rank by omitting the list.

$ dmg pool reintegrate $DAOS_POOL --rank=${rank} --target-idx=${idx1},${idx2},${idx3}

The pool reintegrate command accepts 3 parameters:

  • The label or UUID of the pool that the targets will be reintegrated into.
  • The engine rank of the affected targets.
  • The target indices of the targets to be reintegrated on that engine rank (optional).

When rebuild is triggered it will list the operations and their related engines/targets by their engine rank and target index.

Target (rank 5 idx 0) is down.
Target (rank 5 idx 1) is down.
...
(rank 5 idx 0) is excluded.
(rank 5 idx 1) is excluded.

These should be the same values used when reintegrating the targets.

$ dmg pool reintegrate $DAOS_POOL --rank=5 --target-idx=0,1

Warning

While dmg pool query and list show how many targets are disabled for each pool, there is currently no way to list the targets that have actually been disabled. As a result, it is recommended for now to try to reintegrate all engine ranks one after the other via for i in seq $NR_RANKs; do dmg pool reintegrate --rank=$i; done. This limitation will be addressed in the next release.

Pool Extension

Addition & Space Rebalancing

Full support for online target addition and automatic space rebalancing is planned for a future release and will be documented here once available.

Until then the following command(s) are placeholders and offer limited functionality related to Online Server Addition/Rebalancing operations.

An operator can choose to extend a pool to include ranks not currently in the pool. This will automatically trigger a server rebalance operation where objects within the extended pool will be rebalanced across the new storage.

$ dmg pool extend $DAOS_POOL --ranks=${rank1},${rank2}...

The pool extend command accepts one required parameter which is a comma separated list of engine ranks to include in the pool.

The pool rebalance operation will work most efficiently when the pool is extended to its desired size in a single operation, as opposed to multiple, small extensions.

Resize

Support for quiescent pool resize (changing capacity used on each storage node without adding new ones) is currently not supported and is under consideration.

Pool Catastrophic Recovery

A DAOS pool is instantiated on each target by a set of pmemobj files managed by PMDK and SPDK blobs on SSDs. Tools to verify and repair this persistent data is scheduled for DAOS v2.4 and will be documented here once available.

Meanwhile, PMDK provides a recovery tool (i.e., pmempool check) to verify and possibly repair a pmemobj file. As discussed in the previous section, the rebuild status can be consulted via the pool query and will be expanded with more information.

Recovering Container Ownership

Typically users are expected to manage their containers. However, in the event that a container is orphaned and no users have the privileges to change the ownership, an administrator can transfer ownership of the container to a new user and/or group.

To change the owner user:

$ dmg cont set-owner --pool <UUID> --cont <UUID> --user <owner-user>

To change the owner group:

$ dmg cont set-owner --pool <UUID> --cont <UUID> --group <owner-group>

The user and group names are case sensitive and must be formatted as DAOS ACL user/group principals.

Because this is an administrative action, it does not require the administrator to have any privileges assigned in the container ACL.

Back to top