The storage tiering framework provides iRODS the capability of automatically moving data between any number of identified tiers of storage within a configured tiering group.
To define a storage tiering group, selected storage resources are labeled with metadata which define their place in the group and how long data should reside in that tier before being migrated to the next tier.
The example diagram below shows a configuration with three tiers.
To configure the storage tiering capability the unified_storage_tiering rule engine plugin must be added as a new json object within the rule_engines array of the server_config.json file. By convention the instance_name is the plugin_name with the suffix -instance. Please note that the same instance name must be used across all servers which participate in tiering groups. Data movement is delegated to the server hosting the source resource, and by convention remote rule calls are directed at specific plugin instances.
"rule_engines": [
{
"instance_name": "irods_rule_engine_plugin-unified_storage_tiering-instance",
"plugin_name": "irods_rule_engine_plugin-unified_storage_tiering",
"plugin_specific_configuration": {
}
},
...
]
Tier groups are defined via metadata AVUs attached to the resources which participate in the group.
In iRODS terminology, the attribute is defined by a plugin_specific_configuration named group_attribute (default of irods::storage_tiering::group). The value of the metadata triple is the name of the tier group, and the unit holds the numeric position of the resource within the group. To define a tier group, simply choose a name and apply metadata to the selected root resources of given compositions.
For example:
imeta add -R fast_resc irods::storage_tiering::group example_group_1 0
imeta add -R medium_resc irods::storage_tiering::group example_group_1 1
imeta add -R slow_resc irods::storage_tiering::group example_group_1 2
This example defines three tiers of the group example_group_1 where data will flow from tier 0 to tier 2 as it ages. In this example fast_resc is a single resource, but it could have been set to the root of a resource hierarchy consisting of many resources.
Once a tier group is defined, the age limit for each tier must also be configured via metadata. Once a data object has remained unaccessed on a given tier for more than the configured time, it will be staged to the next tier in the group and then trimmed from the previous tier. This is configured via the default attribute irods::storage_tiering::time (which itself is defined as time_attribute in the plugin_specific_configuration). In order to configure the tiering time, apply an AVU to the resource using the given attribute and a positive numeric value in seconds.
For example, to configure the fast_resc to hold data for only 30 minutes:
imeta add -R fast_resc irods::storage_tiering::time 1800
We can then configure the medium_resc to hold data for 8 hours:
imeta add -R medium_resc irods::storage_tiering::time 28800
Administrators will most likely want the storage tiering policy to execute on a regular basis to check for violating data objects in need of movement.
The simplest way to do this is to add a recurring rule to the delay queue once with irule.
An example rulefile, start_tiering.r:
{
"rule-engine-instance-name": "irods_rule_engine_plugin-unified_storage_tiering-instance",
"rule-engine-operation": "irods_policy_schedule_storage_tiering",
"delay-parameters": "<INST_NAME>irods_rule_engine_plugin-unified_storage_tiering-instance</INST_NAME><PLUSET>1s</PLUSET><EF>60s REPEAT FOR EVER</EF>",
"storage-tier-groups": [
"example_group_1",
"example_group_2"
]
}
INPUT null
OUTPUT ruleExecOut
Update the Execution Frequency (EF) and storage-tier-groups, as appropriate.
Then, execute this rule:
$ irule -r irods_rule_engine_plugin-unified_storage_tiering-instance -F start_tiering.r
The rule will appear in the delay queue:
$ iqstat
id name
10240 {"rule-engine-operation":"irods_policy_storage_tiering","storage-tier-groups":["example_group_1","example_group_2"]}
A number of metadata attributes are used within the storage tiering capability which identify the tier group, the amount of time data may be at rest within the tier, the optional query, etc. These attributes may map to concepts already in use by other names within a given iRODS installation. For that reason we have exposed them as configuration options within the storage tiering plugin_specific_configuration block.
For a default installation the following values are used:
"plugin_specific_configuration": {
"access_time_attribute" : "irods::access_time",
"group_attribute" : "irods::storage_tiering::group",
"time_attribute" : "irods::storage_tiering::time",
"query_attribute" : "irods::storage_tiering::query",
"verification_attribute" : "irods::storage_tiering::verification",
"data_movement_parameters_attribute" : "irods::storage_tiering::restage_delay",
"minimum_restage_tier" : "irods::storage_tiering::minimum_restage_tier",
"preserve_replicas" : "irods::storage_tiering::preserve_replicas",
"object_limit" : "irods::storage_tiering::object_limit",
"default_data_movement_parameters" : "<EF>60s REPEAT UNTIL SUCCESS OR 5 TIMES</EF>",
"minumum_delay_time" : "irods::storage_tiering::minimum_delay_time_in_seconds",
"maximum_delay_time" : "irods::storage_tiering::maximum_delay_time_in_seconds",
"time_check_string" : "TIME_CHECK_STRING",
"data_transfer_log_level" : "LOG_DEBUG"
}
Data movement within a tier group is scheduled asynchronously using the iRODS delayed execution queue, which allows for many jobs to be run simultaneously. In order to prevent the delayed execution server from being overwhelmed a wait time is applied to each job. This time is determined randomly between two separate values configured through metadata. By default the minimum value is 1 second, and the maximum value is 30 seconds. Should a tier within a group expect a high volume of traffic, these values can be adjusted to smaller or larger values.
imeta add -R ufs0 irods::storage_tiering::minimum_delay_time_in_seconds 1
imeta add -R ufs0 irods::storage_tiering::maximum_delay_time_in_seconds 30
When a violating data object is identified for a given source resource, the object is replicated to the next resource in the tier. In order to determine that this operation has succeeded before the source replica is trimmed, the storage tiering plugin provides the ability to perform three methods of verification of the destination replica.
In order of escalating computational complexity, first the system may just rely on the fact that no errors were caught and that the entry is properly registered in the catalog. This is the default behavior, but would also be configured as such:
imeta add -R fast_resc irods::storage_tiering::verification catalog
A more expensive but reliable method is to check the file size of the replica on the destination resource (in this case, when a replica lands on medium_resc, check it):
imeta add -R medium_resc irods::storage_tiering::verification filesystem
And finally, the most computationally intensive but thorough method of verification is computing and comparing checksums. Keep in mind that if no checksum is available for the source replica, such as no checksum was computed on ingest, the plugin will compute one for the source replica first.
imeta add -R slow_resc irods::storage_tiering::verification checksum
After data has been migrated within the system a user may wish to retrieve the data at a future time. When this happens the data is immediately returned to the user, and an asynchronous job is submitted to restage the data to the lowest tier index in the tier group. In the case where an administrator may not wish the data to be returned to the lowest teir, such as when data is automatically ingested, the minimum tier may be indicated with a flag. In this case the storage tiering plugin will restage the data to the indicated tier within the tier group. To configure this option add the following flag to a root resource within the tier group:
imeta add -R medium_resc irods::storage_tiering::minimum_restage_tier true
A tier within a tier group may identify data objects which are in violation by an alternate mechanism beyond the built-in time-based constraint. This allows the data grid administrator to take additional context into account when identifying data objects to migrate.
Data objects which have been labeled via particular metadata, or within a specific collection, owned by a particular user, or belonging to a particular project may be identified through a custom query. The default attribute irods::storage_tiering::query is used to hold this custom query. To configure the custom query, attach the query to the root resource of the tier within the tier group. This query will be used in place of the default time-based query for that tier. For efficiency this example query checks for the existence in the root resource's list of leaves by resource ID. Please note that any custom query must return DATA_NAME, COLL_NAME, USER_NAME, USER_ZONE, DATA_REPL_NUM in that order as it is a convention of this rule engine plugin.
Checking for resources in violating queries is required to prevent erroneous data migrations for replicas on other resources which may represent other tiers in the storage tiering group. This can be done in the manner shown below (DATA_RESC_ID in ('10068', '10069')) or via resource hierarchy (e.g. DATA_RESC_HIER like 'root_resc;%), but the query must filter on resources to correctly identify violating objects.
imeta set -R fast_resc irods::storage_tiering::query "SELECT DATA_NAME, COLL_NAME, USER_NAME, USER_ZONE, DATA_REPL_NUM WHERE META_DATA_ATTR_NAME = 'irods::access_time' AND META_DATA_ATTR_VALUE < 'TIME_CHECK_STRING' AND DATA_RESC_ID IN ('10068', '10069')"
The example above implements the default query. Note that the string TIME_CHECK_STRING is used in place of an actual time. This string will be replaced by the storage tiering framework with the appropriately computed time given the previous parameters.
Any number of queries may be attached in order provide a range of criteria by which data may be tiered, such as user applied metadata. To allow a user to archive their own data via metadata they may tag an object such as archive_object true. The tier may then have a query added to support this.
imeta set -R fast_resc irods::storage_tiering::query "SELECT DATA_NAME, COLL_NAME, USER_NAME, USER_ZONE, DATA_REPL_NUM WHERE META_DATA_ATTR_NAME = 'archive_object' AND META_DATA_ATTR_VALUE = 'true' AND DATA_RESC_ID IN ('10068', '10069')"
Queries may also be provided by using the Specific Query interface within iRODS. The archive object query may be stored by an iRODS administrator as follows.
'iadmin asq "SELECT DISTINCT R_DATA_MAIN.data_name, R_COLL_MAIN.coll_name, R_USER_MAIN.user_name, R_USER_MAIN.zone_name, R_DATA_MAIN.data_repl_num FROM R_DATA_MAIN INNER JOIN R_COLL_MAIN ON R_DATA_MAIN.coll_id = R_COLL_MAIN.coll_id INNER JOIN R_OBJT_ACCESS r_data_access ON R_DATA_MAIN.data_id = r_data_access.object_id INNER JOIN R_OBJT_METAMAP r_data_metamap ON R_DATA_MAIN.data_id = r_data_metamap.object_id INNER JOIN R_META_MAIN r_data_meta_main ON r_data_metamap.meta_id = r_data_meta_main.meta_id INNER JOIN R_USER_MAIN ON r_data_access.user_id = R_USER_MAIN.user_id WHERE r_data_meta_main.meta_attr_name = 'archive_object' AND r_data_meta_main.meta_attr_value = 'true' AND R_DATA_MAIN.resc_id IN ('10068', '10069') ORDER BY R_COLL_MAIN.coll_name, R_DATA_MAIN.data_name, R_DATA_MAIN.data_repl_num" archive_query
At which point the query attached to the root of a storage tier would require the use of a metadata unit of specific:
imeta set -R fast_resc irods::storage_tiering::query archive_query specific
Some users may not wish to trim a replica from a tier when data is migrated, such as to allow data to be archived and also still available on fast storage. To preserve a replica on any given tier, attach the following metadata flag to the root resource.
imeta set -R medium_resc irods::storage_tiering::preserve_replicas true
When working with large sets of data throttling the amount of data migrated at one time can be helpful. In order to limit the results of the violating queries attach the following metadata attribute with the value set as the query limit.
imeta set -R medium_resc irods::storage_tiering::object_limit DESIRED_QUERY_LIMIT
In order to log the transfer of data objects from one tier to the next, set data_transfer_log_level to LOG_NOTICE in the plugin_specific_configuration.
{
"instance_name": "irods_rule_engine_plugin-unified_storage_tiering-instance",
"plugin_name": "irods_rule_engine_plugin-unified_storage_tiering",
"plugin_specific_configuration": {
"data_transfer_log_level" : "LOG_NOTICE"
}
},
There are a few known limitations to the storage tiering plugin which should be noted explicitly for understanding different failure modes which users may experience.
Any given data object should only have replicas in a single tiering group. Stated negatively, a data object should NOT have replicas in multiple tiering groups. The tiering group for a data object is tracked by an AVU that looks like this:
attribute: irods::storage_tiering::group
value: example_group_1
units: 1
The value is the tiering group to which this object belongs. There should only be one AVU with the attribute irods::storage_tiering::group associated with it. Here is an example of the AVUs an object with multiple tiering groups would have:
attribute: irods::storage_tiering::group
value: example_group_1
units: 1
---
attribute: irods::storage_tiering::group
value: example_group_2
units: 2
Notice the different values indicating that the object has replicas in two different tiering groups.
Support for multi-group data objects may become available in the future, but it should be avoided at this time.
Replicating, trimming, and manipulating irods::storage_tiering-namespaced metadata on data objects which are under management in a tiering group using the storage tiering plugin should be done only when necessary. If any metadata is not in the expected state or replicas are not found in their expected resources, unexpected behavior can and will occur as it relates to storage tiering operations.
Resources may belong to multiple tiering groups (e.g. a common archive). It is not recommended for a resource to be tagged with multiple tiers for a given group as the storage tiering plugin assumes that a resource only represents a single tier in any given group. In other words, a resource should not have multiple AVUs like this:
attribute: irods::storage_tiering::group
value: example_group_1
units: 0
---
attribute: irods::storage_tiering::group
value: example_group_1
units: 1
The above AVUs indicate that the resource represents tier 0 AND tier 1 in example_group_1. This should not be done.